Optimal Automatic Stabilizers∗

Alisdair McKay

Boston University

Ricardo Reis

Columbia University and

London School of Economics

June 2016

Abstract

Should the generosity of unemployment benefits and the progressivity of income taxes de-

pend on the presence of business cycles? This paper proposes a tractable model where there is

a role for social insurance against uninsurable shocks to income and unemployment, as well as

inefficient business cycles driven by aggregate shocks through matching frictions and nominal

rigidities. We derive an augmented Baily-Chetty formula showing that the optimal generosity

and progressivity depend on a macroeconomic stabilization term. Using a series of analytical

examples, we show that this term typically pushes for an increase in generosity and progressivity

as long as slack is more responsive to social programs in recessions. A calibration to the U.S.

economy shows that taking concerns for macroeconomic stabilization into account raises the

optimal unemployment benefits replacement rate by 13 percentage points but has a negligible

impact on the optimal progressivity of the income tax. More generally, the role of social insur-

ance programs as automatic stabilizers affects their optimal design.

JEL codes: E62, H21, H30.

Keywords: Counter-cyclical fiscal policy; Redistribution; Distortionary taxes.

∗Contact: amckay@bu.edu and r.a.reis@lse.ac.uk. First draft: February 2015. We are grateful to Ralph Luetticke,Pascal Michaillat, Vincent Sterk, and seminar participants at the SED 2015, the AEA 2016, the Bank of England,Brandeis, Columbia, FRB Atlanta, LSE, Stanford, Stockholm School of Economics, UCSD, the 2016 KonstanzSeminar on Monetary Policy, and the Boston Macro Juniors Meeting for useful comments.

1 Introduction

The usual motivation behind large social welfare programs, like unemployment insurance or pro-

gressive income taxation, is to provide social insurance and engage in redistribution. A large

literature therefore studies the optimal progressivity of income taxes typically by weighing the dis-

incentive effect on individual labor supply and savings against concerns for redistribution and for

insurance against idiosyncratic income shocks.1 In turn, the optimal generosity of unemployment

benefits is often stated in terms of a Baily-Chetty formula, which weighs the moral hazard effect

of unemployment insurance on job search and creation against the social insurance benefits that it

provides.2

For the most part, this literature abstracts from aggregate shocks, so that the optimal generosity

and progressivity do not take into account business cycles. Yet, from their inception, an auxiliary

justification for these social programs was that they were also supposed to automatically stabilize

the business cycle.3 Classic work that did focus on the automatic stabilizers relied on a Keynesian

tradition that ignores the social insurance that these programs provide or their disincentive effects

on employment. More modern work focuses on the positive effects of the automatic stabilizers, but

falls short of computing optimal policies.4

The goal of this paper is to answer two classic questions—How generous should unemployment

benefits be? How progressive should income taxes be?—but taking into account their automatic

stabilizer nature. We present a model in which there is both a welfare role for social insurance as well

as aggregate shocks and inefficient business cycles. Within the model, we introduce unemployment

insurance and progressive income taxes as automatic stabilizers, that is programs that only depend

on the aggregate state of the business cycle indirectly through their dependence on the idiosyncratic

states of the household, which are employment and income.5 We then solve for the ex ante socially

optimal replacement rate of unemployment benefits and progressivity of personal income taxes in

the presence of uninsured income risks, precautionary savings motives, labor market frictions, and

1Mirrlees (1971) and Varian (1980) are classic references, and more recently see Benabou (2002), Conesa andKrueger (2006), Heathcote et al. (2014), Krueger and Ludwig (2013), and Golosov et al. (2016).

2See the classic work by Baily (1978) and Chetty (2006)3Musgrave and Miller (1948) and Auerbach and Feenberg (2000) are classic references, while Blanchard et al.

(2010) is a recent call for more modern work in this topic.4See McKay and Reis (2016) for a recent model, DiMaggio and Kermani (2016) for recent empirical work, and

IMF (2015) for the shortcomings of the older literature.5Landais et al. (2015) and Kekre (2015) instead treat these social programs as active policy choices that vary

directly with the business cycle.

1

nominal rigidities.

Our first main contribution is to provide a formal, theory-grounded definition of an automatic

stabilizer. We show that a business-cycle variant of the Baily-Chetty formula for unemployment

insurance and a similar formula for the optimal choice of progressivity of the tax system are both

augmented by a new macroeconomic stabilization term. This term equals the expectation of the

product of the welfare gain from eliminating economic slack with the elasticity of slack with respect

to the replacement rate or tax progressivity. Even if the economy is efficient on average, economic

fluctuations may lead to more generous unemployment insurance or more progressive income taxes,

relative to standard analyses that ignore the automatic stabilizer properties of these programs.

This terms captures the automatic stabilizer nature of social insurance programs.

The second contribution is to characterize this macroeconomic stabilization term analytically to

understand the different economic mechanisms behind it. Fluctuations in aggregate economic slack,

measured by the unemployment rate, the output gap or the job finding rate, can lead to welfare

losses through four separate channels. First, they may create a wedge between the marginal disu-

tility of hours worked and the social benefit of work. This inefficiency appears in standard models

of inefficient business cycles, and is sometimes described as a result of time-varying markups (Chari

et al., 2007; Galı et al., 2007). Second, when labor markets are tight, more workers are employed

raising production but the cost of recruiting and hiring workers rises. The equilibrium level of un-

employment need not be efficient as hiring and search decisions do not necessarily internalize these

tradeoffs. This is the source of inefficiency common to search models (e.g. Hosios, 1990). Third,

the state of the business cycle alters the extent of uninsurable risk that households face both in

unemployment and income risk. This is the source of welfare costs of business cycles that has

been studied by Storesletten et al. (2001), Krebs (2003, 2007), and De Santis (2007). Finally, with

nominal rigidities, slack affects inflation and the dispersion of relative prices, as emphasized by the

new Keynesian business cycle literature (Woodford, 2010; Gali, 2011). Our measure isolates these

four effects cleanly in terms of separate additive terms in the condition determining the optimal

extent of the social insurance programs.

In turn, the effects of benefits and progressivity of taxes on economic slack depends on their

direct effect on aggregate demand, as well as on the impact of aggregate demand on equilibrium

output. We show that considering macroeconomic stabilization raises the optimal replacement rate

2

of unemployment insurance. This is because, in recessions, economic activity is inefficiently low and

aggregate slack is more responsive to the replacement rate for two reasons. First, there are more

unemployed workers with high marginal propensities to consume receiving the transfers. Second,

the effect of social insurance on precautionary savings motives is larger when there is a greater

risk of unemployment. A similar argument applies to progressive taxation because income risk is

counter-cyclical. Our analysis also incoroprates the effect of other aggregate demand policies and

we show there is little role for fiscal policy to stabilize the business cycle if prices are very flexible

or if monetary policy is very aggressive.

Our third and final contribution is to calculate the optimal automatic stabilizers quantitatively

in the presence of business cycles. We do so by measuring how much more generous is unemployment

insurance and how much more progressive are income taxes in a calibrated economy with aggregate

shocks relative to one where these shocks are turned off. We find a large effect on unemployment

insurance: with business cycles, the optimal unemployment replacement rate rises from 36 to 49

percent. However, the level of tax progressivity has very little stabilizing effect on the business cycle

so the presence of aggregate shocks has almost no effect on the optimal degree of progressivity.

There are large literatures on the three topics that we touch on: business cycle models with

incomplete markets and nominal rigidities, social insurance and public programs, and automatic

stabilizers. Our model of aggregate demand has some of the key features of new Keynesian models

with labor markets (Gali, 2011) but that literature focuses on optimal monetary policy, whereas we

study the optimal design of the social insurance system. Our model of incomplete markets builds on

McKay and Reis (2016), Ravn and Sterk (2013), and Heathcote et al. (2014) to generate a tractable

model of incomplete markets and automatic stabilizers. This simplicity allows us to analytically

express optimality conditions for generosity and progressivity, and to, even in the more general

case, easily solve the model numerically and so be able to search for the optimal policies. Finally,

our paper is part of a surge of work on the interplay of nominal rigidities and precautionary savings,

but this literature has mostly been positive whereas this paper’s focus is on optimal policy.6

On the generosity of unemployment insurance, our work is closest to Landais et al. (2015) and

Kekre (2015). They also generalize the standard Baily-Chetty formula by considering the general

equilibrium effects of unemployment insurance. The main difference is that while they study how

6See Oh and Reis (2012); Guerrieri and Lorenzoni (2011); Auclert (2016); McKay et al. (2016); Kaplan et al.(2016); Werning (2015).

3

benefits should vary over the business cycle, we see how the presence of business cycles affects

the ex ante fixed level of benefits.7 In this sense, our focus is on automatic stabilizers, an ex

ante passive policy, while they consider active stabilization policy. Moreover, our model includes

aggregate uncertainty and we also study income tax progressivity.

On income taxes, our work is closest to Benabou (2002) and Bhandari et al. (2013). Our

dynamic heterogeneous-agent model with progressive income taxes is similar to the one in Benabou

(2002), but our focus is on business cycles, so we complement it with aggregate shocks and nominal

rigidities. Bhandari et al. (2013) are one of the very few studies of optimal income taxes with

aggregate shocks and, like us, they emphasize the interaction between business cycles and the desire

for redistribution.8 However, they do not consider unemployment benefits and restrict themselves

to flat taxes over income. Moreover, they solve for the Ramsey optimal fiscal policy, which adjusts

the tax instruments every period in response to shocks, while we choose the ex ante optimal rules

for generosity and progressivity. This is consistent with our focus on automatic stabilizers, which

are ex ante fiscal systems, rather than counter-cyclical policies.

Finally, this paper is related to the modern study of automatic stabilizers and especially our

earlier work in McKay and Reis (2016). There, we considered how the actual automatic stabilizers

implemented in the US alter the dynamics of the business cycle. Here we are concerned with the

optimal fiscal system as opposed to the observed one.

The paper is structured as follows. Section 2 presents the model, and section 3 discusses its

equilibrium properties. Section 4 derives the macroeconomic stabilization term in the optimality

conditions for the two social programs. Section 5 discusses its qualitative properties, the economic

mechanisms that it depends on, and its likely sign. Section 6 calibrates the model, and quantifies

the effects of the automatic stabilizer effect. Section 7 concludes.

2 The Model

The main ingredients in the model are uninsurable income and employment risks, social insurance

programs, and nominal rigidities so that aggregate demand matters for equilibrium allocations.

The model makes a series of particular assumptions, which we explain in this section, in order to

7See also Mitman and Rabinovich (2011), Jung and Kuester (2015), and Den Haan et al. (2015).8Werning (2007) also studies optimal income taxes with aggregate shocks and social insurance.

4

generate a tractability that is laid out in the next section. Time is discrete and indexed by t.

2.1 Agents and commodities

There are two groups of private agents in the economy: households and firms.

Households are indexed by i in the unit interval, and their type is given by their productivity

αi,t ∈ R+0 and employment status ni,t ∈ 0, 1. Every period, an independently drawn share δ dies,

and is replaced by newborn households with no assets and productivity normalized to αi,t = 1.9

Households derive utility from consumption, ci,t, and publicly provided goods, Gt, and derive

disutility from working for pay, hi,t, searching for work, qi,t, and being unemployed according to

the utility function:

E0

∑t

βt

[log(ci,t)−

h1+γi,t

1 + γ−q1+κi,t

1 + κ+ χ log(Gt)− ξ (1− ni,t)

]. (1)

The parameter β captures the joint discounting effect from time preference and mortality risk,

while ξ is a non-pecuniary cost of being unemployed.10

The final consumption good is provided by a competitive final goods sector in the amount Yt

that sells for price pt. It is produced by combining varieties of goods in a Dixit-Stiglitz aggregator

with elasticity of substitution µ/(µ − 1). Each variety j ∈ [0, 1] is monopolistically provided by

a firm with output yj,t by hiring labor from the households and paying the wage wt per unit of

effective labor.

2.2 Asset markets and social programs

Households can insure against mortality risk by buying an annuity, but they cannot insure against

risks to their individual skill or employment status. The simplest way to capture this market

incompleteness is by assuming that households only trade a single risk-free bond, that has a gross

real return Rt and is in zero net supply. Moreover, we assume that households cannot borrow, so

9The mortality risk allows for a stationary cross-sectional distribution of productivity along with permanent shocksin section 6, but otherwise plays no significant role in the analysis and so will be assumed away in sections 4 and 5.

10If β is pure time discounting, then β ≡ β(1− δ).

5

if ai,t measures their asset holdings:11

ai,t ≥ 0. (2)

The government provides two social insurance programs. The first is a progressive income tax

such that if zi,t is pre-tax income, the after-tax income is λtz1−τi,t . λt ∈ [0, 1] determines the overall

level of taxes, which together with the size of government purchases Gt, will pin down the size of

the government. The object of our study is instead the automatic stabilizer role of the government,

so our focus is on τ ∈ [0, 1]. This determines the progressivity of the tax system. If τ = 0, there

is a flat tax at rate 1 − λt, while if τ = 1 everyone ends up with the same after-tax income. In

between, a higher τ implies a more convex tax function, or a more progressive income tax system.

The second social program is unemployment insurance. A household qualifies as long as it is

unemployed (ni,t = 0) and collects benefits that are paid in proportion to what the unemployed

worker would earn if she were employed. Suppose the worker’s productivity is such that she would

earn pre-tax income zi,t if she were employed, then her after-tax unemployment benefit is bλtz1−τi,t .12

Our focus is on the replacement rate b ∈ [0, 1], with a more generous program understood as having

a higher b.13

Our goal is to characterize the optimal fixed levels of b and τ . Importantly, we consider the ex

ante design problem, so b and τ do not depend on time or on the state of the business cycle. This

corresponds to our focus on their role as automatic stabilizers, programs that can automatically

stabilize the business cycle without policy intervention. We follow the tradition in the literature

on automatic stabilizers that makes a sharp distinction between built-in properties of programs as

opposed to feedback rules or discretionary choices that adjust these programs in response to current

11A standard formulation for asset markets that gives rise to these annuity bonds is the following: A financialintermediary sells claims that pay one unit if the household survives and zero units if the household dies, andsupports these claims by trading a riskless bond with return R. If ai are the annuity holdings of household i, the lawof large numbers implies the intermediary pays out in total (1 − δ)

∫aidi, which is known in advance, and the cost

of the bond position to support it is (1 − δ)∫aidi/R. Because the riskless bond is in zero net supply, then the net

supply of annuities is zero∫aidi = 0, and for the intermediary to make zero profits, Rt = Rt/(1− δ).

12It would be more realistic, but less tractable, to assume that benefits are a proportion of the income the agentearned when she lost her job. But, given the persistence in earnings, both in the data and in our model, our formulationwill not be quantitatively too different from this case. Also, in our notation, it may appear that unemployment benefitsare not subject to the income tax, but this is just the result of a normalization: if they were taxed and the replacementrate was b, then the model would be unchanged and b ≡ b1−τ .

13In our model, focusing on the duration of unemployment benefits instead of the replacement rate would lead tosimilar trade offs, so we refer to b more generally as the generosity of the program.

6

and past information.14

2.3 Key frictions

There are three key frictions in the economy that create the policy trade-offs that we analyze.

2.3.1 Productivity risk

Labor income for an employed household is αitwthit, where αi,t is an idiosyncratic productivity or

skill. The productivity of households evolves as

αi,t+1 = αi,tεi,t+1 with εi,t+1 ∼ F (ε;xt), (3)

and where∫εdF (ε, xt) = 1 for all t, which implies that the average idiosyncratic productivity in

the population is constant and equal to one.15

The distribution of shocks varies over time so that the model generates cyclical changes in

the distribution of earnings risks, as documented by Storesletten et al. (2004) or Guvenen et al.

(2014). We capture this dependence through the variable xt, which captures the aggregate slack

in the economy. A higher xt implies that the economy is tighter, the output gap is positive, or

that the economy is closer to capacity or booming. In the next section we will map xt to concrete

measures of the state of the business cycle like the unemployment rate or the job finding rate. For

concreteness, a simple case that maps to some empirical estimates is to have F (.) be log-normal

with Var(log ε) = σ2(x) and E(log ε) = −0.5σ(x)2.

2.3.2 Employment risk

The second source of risk is employment. We make a strong assumption that unemployment is

distributed i.i.d. across households. Given the high (quarterly) job-finding rates in the US, this is

not such a poor approximation, and it reduces the state space of the model. At the start of the

period, a fraction υ of households loses employment and must search to regain employment. Search

effort qi,t leads to employment with probability Mtqi,t, where Mt is the job-finding rate per unit of

search effort and the probability of resulting in a match is the same for each unit of search effort.

14Perotti (2005) among many others.15Since newborn households have productivity 1, the assumption is that they have average productivity.

7

Therefore, if all households make the same search effort, then aggregate hiring will be υMtqt and

as a result the unemployment rate will be:

ut = υ(1− qtMt). (4)

Each firm begins the period with a mass 1− υ of workers and must post vacancies at a cost to

hire additional workers. As in Blanchard and Galı (2010), the cost per hire is increasing in aggregate

labor market tightness, which is just equal to the ratio of hires to searchers, or the job-finding rate

Mt. The hiring cost per hire is ψ1Mψ2t , denominated in units of final goods where ψ1 and ψ2 are

parameters that govern the level and elasticity of the hiring costs. Since aggregate hires are the

difference between the beginning of period non-employment rate υ and the realized unemployment

rate ut, aggregate hiring costs are:

Jt ≡ ψ1Mψ2t (υ − ut). (5)

We assume a law of large numbers within the firm so the average productivity of hires is 1.

In this model of the labor market, there is a surplus in the employment relationship since, on

one side, firms would have to pay hiring costs to replace the worker and, on the other side, a worker

who rejects a job becomes unemployed and foregoes the opportunity to earn wages this period.

This surplus creates a bargaining set for wages, and there are many alternative models of how

wages are chosen within this set, from Nash bargaining to wage stickiness, as emphasized by Hall

(2005).

We assume a convenient wage rule:

wt = wAt(1− Jt/Yt)xζt . (6)

The real wage per effective unit of labor depends on three variables, aside from a constant w. First,

it increases proportionately with aggregate effective productivity At, as it would in a frictionless

model of the labor market. Second, it falls when aggregate hiring costs are higher, so that some of

these costs are passed from firms to workers. The justification is that when hiring costs rise, the

economy is poorer and this raises labor supply, which the fall in wages exactly offsets. Since these

costs are quantitatively small, in reality and in our calibrations, this assumption has little effect

8

in the predictions of the model but allows us to not have to carry this uninteresting wealth effect

on labor supply throughout the analysis.16 Third, when the labor market is tighter, wages rise,

with an elasticity of ζ. Standard Nash bargaining models lead to a positive dependence between

economic activity and wages, while sticky wage models can be approximated by ζ = 0.

The purpose of this wage rule is to simplify the analysis of the intensive margin of labor supply.

Our analytical results do not depend crucially on the wage rule. Appendix A discusses this at

length, showing that even for a general wage rule, that nests Nash bargaining and many others,

would lead to very similar results. In fact, if labor supply were fixed on the intensive margin, as in

most search models of the labor market, our results would be completely unchanged.

2.3.3 Nominal rigidities

The firm that produces each variety uses the production function yj,t = ηAt hj,tlj,t, where hj,t are

hours per worker and lj,t the workers in the firm, subject to exogenous productivity shocks ηAt .17

The firm’s marginal cost is

wt + ψ1Mψ2t /ht

ηAt.

Marginal costs are the sum of the wage paid per effective unit of labor and the hiring costs that had

to be paid, divided by productivity. Under flexible prices, the firm would set a constant markup, µ,

over marginal cost. The aggregate profits of these firms are distributed among employed workers

in proportion to their skill, which can be thought of as representing bonus payments in a sharing

economy.

However, individual firms cannot choose their actual price to equal this desired price every

period because of nominal rigidities. We consider two separate simple models of nominal rigidities.

In the numerical study of section 6, we assume Calvo (1983) pricing, so that every period a fraction

θ of randomly drawn firms are allowed to change their price, with the remaining 1−θ having to keep

16Moreover, in the special cases of the model studied in section 5, Jt/Yt is a function of xt so this term getsabsorbed by the next term after a redefinition of ζ.

17Given the structure of the labor market, employed workers set their hours taking the hourly wage as given. Weshow below they all make the same choice ht. The firm then chooses how many workers to hire. Marginal cost isthen the cost of increasing the number of workers to produce one more unit of output.

9

their price unchanged from the last period. This leads to the dynamics for inflation πt ≡ pt/pt−1:

πt =

[(1− θ)/

[1− θ

(p∗tpt

)1/(1−µ)]]1−µ

(7)

where p∗t is the price chosen by firms that adjust their price in period t.

In the analytical study of sections 4 and 5 we assume instead a simpler and more transparent

canonical model of nominal rigidities, where every period an i.i.d. fraction θ of firms can set their

prices pj,t = p∗t , while the remaining set their price to equal what they expected their optimal price

would be: pj,t = Et−1 p∗t . Mankiw and Reis (2010) show that most of the qualitative insights from

New Keynesian economics can be captured by this simple sticky-information formulation.

2.4 Other government policy

Aside from the two social programs that are the focus of our study, the government also chooses

policies for nominal interest rates, government purchases, and the public debt. Starting with the

first, we assume a standard Taylor rule for nominal interest rates It:

It = Iπωπt xωxt ηIt , (8)

where ωπ > 1 and ωx ≥ 0. The exogenous ηIt represent shocks to monetary policy.18

Turning to the second, government purchases follow the Samuelson (1954) rule such that, absent

ηGt shocks, the marginal utility benefit of public goods offsets the marginal utility loss from diverting

goods from private consumption:

Gt = χCtηGt . (9)

Finally, we assume that the government runs a balanced budget:

Gt =

∫ni,t

(zi,t − λtz1−τi,t

)− (1− ni,t) bλtz1−τi,t di, (10)

where zi,t is the income of household i should they be employed. This strong assumption deserves

some explanation. It is well known, at least since Aiyagari and McGrattan (1998), that in an

18As usual, the real and nominal interest rates are linked by the Fisher equation Rt = It/Et [πt+1].

10

incomplete markets economy like ours, changes in the supply of safe assets will affect the ability

to accumulate in precautionary savings. Deficits or surpluses may stabilize the business cycle by

changing the cost of self-insurance. In the same way that we abstracted above from the stabilizing

properties of changes in government purchases, this lets us likewise abstract from the stabilizing

property of public debt, in order to focus on our two social programs.

In previous work (McKay and Reis, 2016), we found that allowing for deficits and public debt had

little effect on the effectiveness of stabilizers. This is because, in order to match the concentration

of wealth in the data, almost all of the public debt is held by richer households who are already

close to fully self insured. Still, in previous versions of this paper, we investigated this further by

assuming instead that there are public deficits, but financed by borrowing from abroad. As long

as changes in b and τ do not affect the amount of debt issued, all of our results are unchanged.

3 Equilibrium and the role of policy

Our model combines idiosyncratic risk, incomplete markets, and nominal rigidities, and yet it is

structured so as to be tractable enough to investigate optimal policy. This section highlights how

our assumptions, with their virtues and limitations, lead to this tractability. We also highlight the

role for social insurance policy in the economy, as well as the distortions it creates.

3.1 Inequality and heterogeneity

The following result follows from the particular assumptions we made on the decision problems of

different agents and plays a crucial role in simplifying the analysis:

Lemma 1. All households choose the same asset holdings, hours worked, and search effort, so

ai,t = 0, hi,t = ht, and qi,t = qt for all i.

To prove this result, note that the decision problem of a household searching for a job at the

start of the period is:

V s(a, α,S) = maxq

MqV (a, α, 1,S) + (1−Mq)V (a, α, 0,S)− q1+κ

1 + κ

, (11)

where we used S to denote the collection of aggregate states. The decision problem of the household

11

at the end of the period is:

V (a, α, n,S) = maxc,h,a′≥0

log c− h1+γ

1 + γ+ χ log(G)− ξ(1− n)+

βE[(1− υ)V (a′, α′, 1,S ′) + υV s(a′, α′,S ′)

], (12)

subject to: a′ + c = Ra+ λ (n+ (1− n)b) [α(wh+ d)]1−τ . (13)

Starting with asset holdings, since no agent can borrow and bonds are in zero net supply, then

it must be that ai,t = 0 for all i in equilibrium because there is no gross supply of bonds for savers

to own, a result also used by Krusell et al. (2011) and Ravn and Sterk (2013). Turning to hours

worked, the intra-temporal labor supply condition for an employed household is

ci,thγi,t = (1− τ)λtz

−τi,t wtαi,t, (14)

where the left-hand side is the marginal rate of substitution between consumption and leisure, and

the right-hand side is the after-tax return to working an extra hour to raise income zi,t. More

productive agents want to work more. However, they are also richer and want to consume more.

The combination of our preferences and the budget constraint imply that these two effects exactly

cancel out so that in equilibrium all employed households work the same hours:

hγt =(1− τ)wtwtht + dt

, (15)

where dt is aggregate dividends per employed worker.19

Finally, the optimality condition for search effort is:

qκi,t = Mt [V (ai,t, αi,t, 1,S)− V (ai,t, αi,t, 0,S)] . (16)

Intuitively, the household equates the marginal disutility of searching on the left-hand side to

the expected benefit of finding a job on the right-hand side, which is the product of the job-

finding probability Mt and the increase in value of becoming employed. Appendix B.1 shows

that this increase in value is independent of αi,t. The key assumption that ensures this is that

19To derive this, substitute zi,t = αi,t(wthi,t + dt) and ci,t = λtz1−τi,t into (14).

12

unemployment benefits are indexed to income zi,t so the after-tax income with and without a job

scales with idiosyncratic productivity in the same way. This then implies that qi,t is the same for

all households.

The lemma clearly limits the scope of our study. We cannot speak to the effect of policy on asset

holdings, and differences in labor supply are reduced to having a job or not, which ignores diversity

in part-time jobs and overtime. At the same time, it has the substantial payoff of implying that S

contains only aggregate variables, so we do not need to keep track of cross-sectional distributions to

characterize an equilibrium. Thus, our model can be studied analytically and numerical solutions

are easy to compute. Moreover, arguably the social programs that we study are more concerned

with income, rather than wealth inequality, and the vast majority of studies of the automatic

stabilizers also ignores any direct effects of wealth inequality (as opposed to income inequality) on

the business cycle.

In our model, there is a rich distribution of income and consumption driven by heterogeneity in

employment status ni,t and skill ai,t. In section 6, we are able to fit the more prominent features of

income inequality in the United States by parameterizing the distribution F (ε, x). Moreover, in our

model, there is a rich distribution of individual prices and output across firms, (pj,t, yj,t), driven by

nominal rigidities. And finally, the exogenous aggregate shocks to productivity, monetary policy,

and government purchases, (ηAt , ηIt , η

Gt ), affect all of these distributions, which therefore vary over

time and over the business cycle. In spite of the simplifications and their limitations, our model

still admits a rich amount of inequality and heterogeneity.

3.2 Quasi-aggregation and consumption

Define ct as the consumption of the average-skilled (αi,t = 1), employed agent. This is related to

aggregate consumption, Ct, according to (see Appendix B.2):

ct =Ct

Ei[α1−τi,t

](1− ut + utb)

. (17)

Funding higher replacement rates requires larger taxes on those employed, so it reduces their

consumption. Likewise, the amount of revenue raised by the progressive tax system depends on

the distribution of income as summarized by Ei[α1−τi,t

]. More dispersed incomes generate higher

13

revenues and allow for lower taxes for a given level of income.

The next property that simplifies our model is proven in Appendix B.2.

Lemma 2. Aggregate consumption dynamics can be computed from

1

ct= βRt Et

1

ct+1Qt+1

, (18)

with: Qt+1 ≡[(1− ut+1) + ut+1b

−1]E [ε−(1−τ)i,t+1

](19)

and equation (17).

Without uncertainty on productivity or unemployment, Qt+1 = 1, and this would be a standard

Euler equation from intertemporal choice stating that expected consumption growth is inversely

related to the product of the discount factor and the real interest rate.

The variable Qt+1 captures how heterogeneity affects aggregate consumption dynamics through

precautionary savings motives. The more uncertain is income, the larger is Qt+1 and so the larger

are savings motives leading to steeper consumption growth. A more generous unemployment in-

surance system and a more progressive income tax lower the dispersion of after-tax income growth

and reduce the effect of this Qt+1 term.

3.3 Policy distortions and redistribution over the business cycle

Social policies not only affect aggregate consumption, but also all individual choices in the economy,

introducing both distortions and redistribution.

Combining the optimality condition for hours with the wage rule gives (see Appendix B.3):

ht = [w(1− τ)]1

1+γ xζ

1+γ

t . (20)

A more progressive income tax lowers hours worked by increasing the ratio of the marginal tax rate

to the average tax rate.

Moving to search effort, one can show that (see Appendix B.3)

qκt = Mt

[ξ − h1+γt

1 + γ− log(b)

]. (21)

14

This states that the marginal disutility of searching for a job is equal to the probability of finding a

job times the increase in utility of having a job. This increase is equal to the difference between the

non-pecuniary pain from being unemployed and the disutility of working, minus the loss in utility

units of losing the unemployment benefits. More generous benefits therefore lower search effort.

Intuitively, they lower the value of finding a job, so less effort is expended looking for one.

The distribution of consumption in the economy is given by a relatively simple expression:

ci,t =[α1−τi,t (ni,t + (1− ni,t)b)

]ct (22)

The expression in brackets shows that more productive and employed households consume more,

as expected. Combined with ct, this formula also shows how social policies redistribute income and

equalize consumption. A higher b requires larger contributions from all households, lowering ct, but

only increases the term in brackets for unemployed households. Therefore, it raises the consumption

of the unemployed relative to the employed. In turn, a higher τ lowers the cross-sectional dispersion

of consumption because it reduces the income of the rich more than that of the poor. The state of

the business cycle affects the extent of the redistribution by driving both unemployment and the

cross-sectional distribution of productivity risk.

Finally, social programs also affect price dispersion and inflation. Recalling thatAt ≡ Yt/[ht∫ljtdj],

average aggregate labor productivity, then integrating over the individual production functions and

using the demand for each variety it follows immediately that At = ηAt /St where the new variable

is price dispersion:

St ≡∫

(pt(j)/pt)µ/(1−µ) dj ≥ 1 (23)

= (1− θ)St−1π−µ/(1−µ)t + θ

(p∗tpt

)µ/(1−µ)if Calvo, (24)

=

(p∗tpt

)µ/(1−µ) [θ + (1− θ)

(Et−1p∗tp∗t

)µ/(1−µ)]if sticky information. (25)

Nominal rigidities lead otherwise identical firms to charge different prices, and this relative-price

dispersion lowers productivity and output in the economy. The social insurance system will alter the

dynamics of aggregate demand leading to different dynamics for nominal marginal costs, inflation,

and price dispersion.

15

3.4 Slack and equilibrium

The missing ingredient to close the model is a concrete definition of how to measure economic slack

xt.

For the analytical results in the next two sections, we will take a convenient assumption:

xt = Mt. (26)

That is, we measure the state of the business cycle by the tightness of the labor market, as captured

by the job-finding rate. This is not such a strong assumption since, in the model, (ht, qt) are

functions of only Mt and parameters, as we can see in equations (20) and (21). Moreover, in

the special cases considered in sections 4 and 5, the unemployment rate and the output gap (the

difference between actual output and that which arises with flexible prices) are also functions of

Mt as the single endogenous variable. Finally, when we take the model to the data in section 6,

we find that in simulations, using instead one minus the unemployment rate as the measure of

slack leads to essentially identical results. We assume equation (26) because it makes the analytical

derivations in the next two sections more transparent, allowing us to carry fewer cross-terms that

are of little interest. Most of the results would extend easily to other measures of slack, like the

unemployment rate or hours worked, but with longer and more involved algebraic expressions.

An aggregate equilibrium in our economy is then a solution for 18 endogenous variables together

with the exogenous processes ηAt , ηGt , and ηIt . Appendix B.4 lays out the entire system of equations

that defines this equilibrium.

4 Optimal policy and insurance versus incentives

All agents in our economy are identical ex ante, making it natural to take as the target of policy

the utilitarian social welfare function. Using equation (22) and integrating the utility function in

equation (1) gives the objective function for policy E0∑∞

t=0 βtWt, where period-welfare is:

Wt = Ei log(α1−τi,t

)− log

(Ei[α1−τi,t

])+ ut log b− log (1− ut + utb)

+ log(Ct)− (1− ut)h1+γt

1 + γ− υ q

1+κt

1 + κ+ χ log(Gt)− ξut. (27)

16

The first line shows how inequality affects social welfare. Productivity differences and unemploy-

ment introduce costly idiosyncratic risk, which is attenuated by the social insurance policies. The

second line captures the usual effect of aggregates on welfare. While these would be the terms that

would survive if there were complete insurance markets, recall that the incompleteness of markets

also affects the evolution of aggregates, as we explained in the previous section.

The policy problem is then to pick b and τ to maximize equation (27) subject to the equilibrium

conditions, at date 0 once and for all. As already discussed, in this and the next section, we make

the following simplifications on the general problem: (i) log-normal productivity shocks, (ii) no

mortality, (iii) sticky information, and (iv) no government spending shocks.20

4.1 Optimal unemployment insurance

Appendix C derives the following optimality condition for b:

Proposition 1. The optimal choice of the generosity of unemployment insurance b satisfies:

E0

∞∑t=0

βt

ut

(1

b− 1

)∂ log (bct)

∂ log b

∣∣∣∣x,q

+d log ctd log ut

∂ log ut∂b

∣∣∣∣x

+dWt

dxt

dxtdb

= 0. (28)

Equation (28) is closely related to the Baily-Chetty formula for optimal unemployment insur-

ance. The first term captures the social insurance value of changing the replacement rate. It is

equal to the percentage difference between the marginal utility of unemployed and employed agents

times the elasticity of the consumption of the unemployed with respect to the benefit. If unem-

ployment came with no differences in consumption, this term would be zero, and likewise if giving

higher benefits to the unemployed had no effect on their consumption. But as long as employed

agents consume more, and raising benefits closes some of the consumption gap, then this term will

be positive and call for higher unemployment benefits.

The second term gives the moral hazard cost of unemployment insurance. Is is equal to the

product of the elasticity of the consumption of the employed with respect to the unemployment

rate, which is negative, and the elasticity of the unemployment rate with respect to the benefit that

20To be clear, none of these assumptions are essential: relaxing (i) would lead to similar expressions with expecta-tions against F (.) in place of σ2(.), relaxing (ii) would result in more complicated expressions for the effect of skillrisk on welfare without any qualitative change in the results, substituting (iii) for sticky prices would require inte-grating the effects of policy on St over time, and relaxing (iv) would lead to an additional term in all the expressionsequal to the difference between the marginal utility of public expenditures and the resource cost of financing thoseexpenditures.

17

arises out of reduced search effort. Higher replacement rates induce agents to search less, which

raises equilibrium unemployment, and leads to higher taxes to finance benefits.

In the absence of general equilibrium effects, these would be the only two terms, and they capture

the standard trade-off between insurance and incentives in the literature averaged across states

and time. With business cycles and general equilibrium effects, there is an extra macroeconomic

stabilization term. The larger this term is, the more generous optimal unemployment benefits

should be. We explain this shortly, but first, we turn to the income tax.

4.2 Optimal progressivity of the income tax

Appendix C shows the following:

Proposition 2. The optimal progressivity of the tax system τ satisfies:

E0

∞∑t=0

βt[

β(1− τ)

(1− β)

]σ2(xt)−

(AtCt− hγt

)ht(1− ut)

(1− τ)(1 + γ)+d log ctd log ut

∂ log ut∂τ

∣∣∣∣x

+dWt

dxt

dxtdτ

= 0.

(29)

The first three terms again capture the familiar trade-off between insurance and incentives.

The first term gives the welfare benefits of reducing the dispersion in after-tax incomes, which is

increasing in the extent of pre-tax inequality as reflected by σ2(xt). The second and third term give

the incentive costs of raising progressivity. The second term is the labor wedge, the gap between

the marginal product of labor and the marginal disutility of labor. More progressive taxes raise the

wedge by discouraging labor supply, as explained earlier. The third term reflects the effect of the

tax system on the unemployment rate taking slack as given. The tax system affects the relative

rewards to being employed and therefore alters household search effort and the unemployment rate.

Finally, the fourth term captures the concern for macroeconomic stabilization in a very similar

way to the term for unemployment benefits. A larger stabilization term in (29) justifies a larger

labor wedge and therefore a more progressive tax.

4.3 The macroeconomic stabilization term

The two previous propositions clearly isolate the automatic-stabilizing role of the social insurance

programs in a single term. It equals the product of the welfare benefit of changing slack and the

18

response of slack to policy. If business cycles are efficient, the macroeconomic stabilization term

is zero. That is, if the economy is always at an efficient level of slack, so that dWt/dxt = 0, then

there is no reason to take macroeconomic stabilization into account when designing the stabilizers.

Intuitively, the business cycle is of no concern for policymakers in this case.

Even if business cycles are efficient on average though, the automatic stabilizers can still play

a role. This is because:

E0

∞∑t=0

βtdWt

dxt

dxtdb

=∞∑t=0

βtE0

[dWt

dxt

]E0

[dxtdb

]+ Cov

[dWt

dxt,dxtdb

],

so that even if E0 [dWt/dxt] = 0, a positive covariance term would still imply a positive aggregate

stabilization term and an increase in benefits (or more progressive taxes). Our model therefore

provides a sharp definition of the the hallmark of a social policy that serves as an automatic

stabilizer: it stimulates the economy more in recessions, when slack is inefficiently high. The

stronger this effect, the larger the program should be. In the next section, we discuss the sign of

this covariance and what affects it.

5 Inspecting the macroeconomic stabilization term

Understanding the automatic stabilizer nature of social program requires understanding separately

the effect of slack on welfare, dWt/dxt, and the effect of the social policies on slack, dxt/db and

dxt/dτ . Instead of trying to measure the covariance between these two unobservables in the data,

a daunting task, we proceed by characterizing their structural determinants instead in terms of

familiar economic channels that have been measure elsewhere.

5.1 Slack and welfare

There are five separate channels through which the business cycle may be inefficient in our model,

characterized in the following result:

19

Proposition 3. The effect of macroeconomic slack on welfare can be decomposed into:

dWt

dxt= (1− ut)

[AtCt− hγt

]dhtdxt︸ ︷︷ ︸

labor-wedge

− YtCtSt

dStdxt︸ ︷︷ ︸

price-dispersion

+1

Ct

dCtdut

dutdxt− 1

Ct

∂Jt∂xt

∣∣∣∣u︸ ︷︷ ︸

Hosios

(30)

−

(ξ − log b− h1+γt

1 + γ

)∂ut∂xt

∣∣∣∣q

+1− b

1− ut + utb

dutdxt︸ ︷︷ ︸

unemployment-risk

− β(1− τ)2

2(1− β)

dσ2(xt)

dxt︸ ︷︷ ︸income-risk

The first term captures the effect of the labor wedge or markups. In the economy, At/Ct is

the marginal product of an extra hour worked in utility units, while hγt is the marginal disutility

of working. If the first exceeds the second, the economy is underproducing, and increasing hours

worked would raise welfare.

The second term captures the effect of slack on price dispersion. Because of nominal rigidities,

aggregate shocks will lead to price dispersion. In that case, changes in aggregate slack will affect

inflation, via the Phillips curve, and so price dispersion. This is the conventional channel in new

Keynesian models through which the output gap affects inflation and its welfare costs.

The third and fourth terms capture the standard Hosios (1990) trade-off of hiring more workers.

On the one hand, the extra hire lowers unemployment and raises consumption. On the other hand, it

increases hiring costs. If hiring is efficient, so the Hosios condition holds, then (dCt/dut)(dut/dxt) =

∂J/∂xt at all dates, but otherwise changes in slack will affect hirings, unemployment and welfare.

The terms in the second line of equation (30) fix aggregate consumption and focus on inequality

and its effect on welfare. If the extent of income risk is cyclical, which the literature since Storeslet-

ten et al. (2004) has extensively demonstrated, then raising economic activity reduces income risk

and so raises welfare. In our model, there is both unemployment and income risk, so this works

through two channels.

The fourth and fifth term capture the effect of slack on unemployment risk. For a given aggregate

consumption, more unemployment has two effects on welfare. First there are more unemployed

who consume a lower amount. The term ξ − log b− h1+γt /(1 + γ) is the utility loss from becoming

unemployed. Second, those who are employed consume a larger share (dividing the pie among fewer

employed people). These are the two effects of unemployment risk. The sixth and final term shows

that cutting slack also lowers the variance of skill shocks, which lowers income inequality.

20

5.2 Three special cases

To better understand these different channels of stabilization, and link them to the literature before

us, we consider three special cases that correspond to familiar models of fluctuations.

5.2.1 Frictional unemployment

Consider the special case where prices are flexible (θ = 1), there is no productivity risk (σ2 = 0),

and labor supply does not vary on the intensive margin because hours worked are constant (γ =∞).

The only source of inequality is then unemployment, which in our model becomes a result of the

search and matching paradigm. Therefore, equation (30) becomes:

dWt

dxt=

1

Ct

dCtdut

dutdxt− 1

Ct

∂Jt∂xt

∣∣∣∣u

−

(ξ − log b− h1+γt

1 + γ

)∂ut∂xt

∣∣∣∣q

+1− b

1− ut + utb

dutdxt

. (31)

as only the Hosios effect and the unemployment risk are now present.

In this special case, our model captures the main effects in Landais et al. (2015). They discuss

the macroeconomic effects of unemployment benefits from the perspective of their effect on labor

market tightness by changing the worker’s bargaining position and wages on the one hand and, on

the other hand, their impact on dissuading search effort.

5.2.2 Real Business Cycle effects

Next, we consider the case of flexible prices (θ = 1), constant search effort (κ =∞), and exogenous

job finding (Mt exogenous).21 With nominal rigidities and search removed, what is left is the labor

wedge and the effect of cyclical income risk on welfare, so equation (30) simplifies to

dWt

dxt= (1− ut)

[AtCt− hγt

]dhtdxt− β

1− β(1− τ)2

d

dxt

σ2(xt)

2. (32)

In this case, our paper fits into the standard analysis of business cycles in Chari et al. (2007)

through the first term, and into the study the costs of business cycles due to income inequality

emphasized by Krebs (2003) through the second term.

21When Mt is constant we need to define slack differently from xt = Mt. In this case, the role of xt is to changethe wage and change labor supply on the intensive margin. The wage will need to adjust to clear the labor marketas in the three-equation New Keynesian model and then the wage rule, equation (6), becomes the definition of xt.

21

5.2.3 Aggregate-demand effects

Traditionally, the literature on automatic stabilizers has focussed on aggregate demand effects

following a Keynesian tradition. When there is no productivity risk (σ2 = 0), job search effort

is constant (κ = ∞) and the labor market’s matching frictions are constant (Mt is constant),

equation(30) simplifies to:

dWt

dxt= (1− ut)

[AtCt− hγt

]dhtdxt− YtCtSt

dStdxt

,

so only the markup effects are present, both through the labor wedge and through price dispersion.

Appendix D.2 shows that a second-order approximation of Wt around the flexible-price, socially-

efficient level of aggregate output Y ∗t and consumption C∗t transforms this expression into

dWt

dxt=

(Y ∗tC∗t

)[(1

C∗t+

γ

Y ∗t

)(Y ∗t − YtY ∗t

)dYtdxt

+

(1− θθ

µ

µ− 1

)(Et−1pt − pt

Et−1pt

)dptdxt

],

In this case, our model fits into the new Keynesian framework with unemployment developed in

Blanchard and Galı (2010) or Gali (2011). Raising slack affects the output gap and the price

level, through the Phillips curve, and this affects welfare through the two conventional terms in the

expression. The first is the effect on the output gap, and the second the effect on surprise inflation.

These are the two sources of welfare costs in this economy.

5.3 Social programs and slack

We now turn attention to the second component of the macroeconomic stabilization term, either

dxt/db in the case of unemployment benefits or dxt/dτ in the case of tax progressivity. We make a

few extra simplifying assumptions to obtain analytical expressions that are easy to interpret. First,

we assume that aggregate shocks only occur at date 0 and there is no aggregate uncertainty after

that, so that the analysis can be contained to a single period. Second, we assume that household

search effort is exogenous and constant (κ =∞) as in sections 5.2.2 and 5.2.3, since the role of job

search is well described by Landais et al. (2015).

22

x0

Y0

ADAS

Policy

Figure 1: Determination of x0.

5.3.1 An AD-AS interpretation

With these simplifications, the analysis of the model becomes static, and the household decisions

are reduced to consumption and hours worked. Appendix D.3 shows that combining equilibrium

in the labor market with the resource constraint and production functions gives an upward sloping

relationship between slack x and output Y , which we will call the “aggregate supply curve”, for

lack of a better term. Likewise combining the aggregate Euler equation, the monetary rule and the

Phillips curve gives an “aggregate demand curve.” Both are plotted in figure 1, so that equilibrium

is at the intersection of the two curves.

The impact of the social policies on equilibrium slack depends on two features of the diagram.

First, how much does a change in policy horizontally shift the AD curve? A bigger shift in AD will

lead to a larger effect on slack. Second, what are the slopes of the two curves? If both the AD and

the AS are steeper, then a given horizontal shift in the AD brought about by the policies leads to

a larger change in slack at the new equilibrium. We discuss shifts and slopes separately.

5.3.2 Unemployment benefits and slack

Appendix D.4 proves the following:

23

Proposition 4. Under the assumptions of section 5.3:

d log x0d log b

= Λ−1[

u0b

1− u0 + u0b+

u1b−1

1− u1(1− b−1)− u1b

1− u1(1− b)

](33)

where Λ is defined below in lemma 3.

There are two direct effects of raising unemployment benefits on economic slack. The first

shows the usual logic of the unemployment insurance system as an automatic stabilizer based on

redistribution: aggregate demand responds more strongly to benefits if there are more unemployed

workers. This effect arises because the unemployed have a high marginal propensity to consume so

the extra benefits lead to an increase in demand.

Second, there is an additional effect coming from expected marginal utility in the future, which

depends on the unemployment rate in the next period (t = 1). Expected marginal utility is deter-

mined by two components. The first is the social insurance that UI provides, lowering uncertainty

and precautionary savings and so pushing up aggregate demand today. The second is the higher

taxes in order to finance the higher benefits, which reduce consumption. For unemployment rates

less than 50% the former effect dominates and the sum of these terms is increasing in the unem-

ployment rate u1 (see appendix D.5).

These two effects, captured by the expression in square brackets, shift the AD curve rightwards.

As the unemployment rate increases in recessions, both of these effects point towards a counter-

cyclical elasticity. The overall effect on tightness is then tempered by the slopes of AD and AS,

through the term Λ. We explain this effect below, but since it is common to the effect of more

progressive taxes, we turn to that first.

5.3.3 Tax progressivity and slack

Appendix D.6 proves:

Proposition 5. Under the assumptions of section 5.3:

d log x0d log τ

= Λ−1[2σ2ε (x0) (1− τ) τ − ∂ logR0

∂ log τ

∣∣∣∣x

+∂ logS0∂ log τ

∣∣∣∣x

+d log Y1d log x1

d log x1d log τ

](34)

where Λ is defined below in lemma 3.

24

The first term in between square brackets reflects the effect of social insurance on aggregate

demand. When households face uninsurable skill risk, a progressive tax system will raise aggregate

demand by reducing the precautionary savings motive. This term will be counter-cyclical if risk

increases in a recession as has been documented by Guvenen et al. (2014).

The remaining terms inside the brackets reflect the effect of a change in τ on marginal cost

holding xt fixed. These terms arise because hiring costs are spread over the hours of the worker so

that even at given levels of wt and Mt, a more progressive tax would lower hours worked per worker

and raise hiring costs per hour worked. If labor supply were fixed on the intensive margin, these

terms would not be present. With a choice of hours of work, we see that higher marginal costs put

upward pressure on prices which is reflected in real interest rates via the monetary policy rule and

in price dispersion. For similar reasons, an increase in τ puts upward pressure on marginal cost in

period 1. As prices are flexible in period 1 this force leads to a reduction in labor market tightness

in period 1 so that marginal cost is again equal to the inverse of the desired markup. Because these

effects all operate through rescaling the hiring costs, which are themselves small, they are not likely

to be important quantitatively.

5.3.4 Slopes of AS and AD

Finally, we turn to Λ, which reflects the slopes of the AS and AD curves. As the previous two

propositions showed, a larger Λ attenuates the effect of the social policies on slack, because it makes

both AS and AD flatter. The next lemma, proven in appendix D.7, describes what determines it:

Lemma 3. Under the assumptions of section 5.3:

Λ =d logR0

d log x0+ (1− τ)2σ2ε (x0)

d log σ2ε (x0)

d log x0+

1− b1− u0 + u0b

u0d log u0d log x0

+d log h0d log x0

+d log(1− u0)d log x0

+d log(1− J0/Y0)

d log x0− d logS0d log x0

(35)

The first line has the three terms that affect the slope of the AD curve. The first effect involves

the real interest rate. A booming economy leads to higher nominal interest rates, both directly

via the Taylor rule and indirectly via higher inflation. With nominal rigidities, this raises the real

interest rate, which dampens the effectiveness of any policy on equilibrium slack. In other words, a

more aggressive monetary policy rule (or more flexible prices) makes AD flatter and so attenuates

25

the effectiveness of social policies. An extreme example of this is when the economy is at the zero

lower bound, which magnifies the effectiveness of the automatic stabilizers.22

The second term refers to income risk. If risk is counter-cyclical, this term is negative, so it

makes social programs more powerful in affecting slack. The reason is that there is a destabilizing

precautionary savings motive that amplifies demand shocks. In response to a reduction in aggre-

gate demand, labor market tightness falls, leading to an increase in risk and an increase in the

precautionary savings motive and so a further reduction in aggregate demand. These reinforcing

effects make the AD curve steeper so that social policies become more effective.23

The third term reflects the impact of economic expansions on the number of employed house-

holds, who consume more than unemployed households. Therefore, aggregate demand rises as

employment rises in a tighter labor market. This consumption multiplier makes the AD curve

steeper and increases the effectiveness of social programs.

The second line in the lemma has the four terms that affect the slope of the AS. Increasing

slack raises hours worked or employment, this makes the AS flatter as output increases by relatively

more, so it raises Λ and attenuates the effect of social programs. This occurs net of hiring costs,

since the fact that they increase with slack works in the opposite direction. Finally, if in a booming

economy the efficiency loss from price dispersion increases, so S increases, then the AS is likewise

steeper and Λ is lower.

5.4 Summary and likely sign

To summarize, there are two main channels through which unemployment benefits or income tax

progressivity raise aggregate demand and thereby eliminate slack. These channels are redistribution

and social insurance, and both are increasing in the unemployment rate. As unemployment and

income risks are counter-cyclical, these forces push for counter-cyclical elasticities of slack to the

social programs, since they dampen the counter-cyclical fluctuations in the precautionary savings

motive. If business cycles are inefficient in the sense that tightness is inefficiently low in a recession,

then we expect a positive covariance between dWt/dxt and the elasticities of tightness with respect

to policy. This positive covariance implies a positive aggregate stabilization term and more generous

22 McKay and Reis (2016) and Kekre (2015) show that automatic stabilizers and unemployment benefits, respec-tively, have stronger stimulating effects when the economy is at the zero lower bound.

23Similar reinforcing dynamics arise out of unemployment risk in Ravn and Sterk (2013), Den Haan et al. (2015),and Heathcote and Perri (2015).

26

unemployment benefits and a more progressive tax system even if the business cycle is efficient on

average.

6 Quantitative analysis

We have shown that business cycles lead to a macroeconomic stabilization term that affects the

optimal generosity of unemployment insurance and the progressivity of income taxes, and that this

term likely makes these programs more generous and progressive, respectively, through multiple

channels. We now ask whether these effects are quantitatively significant. To do so, we solve and

calibrate the model in sections 2 and 3, without the simplifications made in sections 4 to 5 used for

analytical characterizations.

6.1 Calibration and solution of the model

In our baseline case, we define slack as xt = (1−ut)/(1−u) where u is the steady state unemployment

rate, instead of the job-finding rate as before. This makes the calibration more transparent since the

unemployment rate is a more commonly used indicator of labor market conditions.24 For similar

reasons, we use the Calvo model of price stickiness. We solve the model using global methods as

described in Appendix E.2, so that we can accurately compute social welfare, assuming that the

economy starts at date 0 at the deterministic steady state. We then numerically search for the

values of b and τ that maximize the social welfare function, and compare these with the maximal

values in a counterfactual economy without aggregate shocks, but otherwise identical.

Table 1 shows the calibration of the model, dividing the parameters into different groups. The

first group has parameters that we set ex ante to standard choices in the literature. Only the last

one deserves some explanation. ψ2 is the elasticity of hiring costs with respect to labor market

tightness, and we set it at 1 as in Blanchard and Galı (2010), in order to be consistent with an

elasticity of the matching function with respect to unemployment of 0.5 as suggested by Petrongolo

and Pissarides (2001).

Panel B contains parameters individually calibrated to hit some time-series moments. For the

preference for public goods, we target the observed average ratio of government purchases to GDP

24Quantitatively though, it makes no difference as in our baseline specification there is an almost perfect relationshipbetween the job finding rate and this measure of slack so with an appropriate recalibration of the model we couldtreat the job-finding rate as the measure of slack.

27

Symbol Value Description Target

Panel A. Parameters chosen ex ante

δ 1/200 Mortality rate 50 year expected lifetimeµµ−1 6 Elasticity of substitution Basu and Fernald (1997)

θ 1/3.5 Prob. of price reset Klenow and Malin (2010)1/γ 1/2 Frisch elasticity Chetty (2012)ψ2 1 Elasticity of hiring cost Blanchard and Galı (2010)

Panel B. Parameters individually calibrated

χ 0.262 Preference for public goods G/Y = 0.207ωπ 1.66 Mon. pol. response to π Estimated Taylor ruleωu 0.133 Mon. pol. response to u Estimated Taylor ruleυ 0.153 Job separation rate Average value

F (ε, .) mix-normals Skill-risk process Guvenen and McKay (2016)

Panel C. Parameters jointly calibrated to steady-state moments

β 0.974 Discount factor 3% annual real interest ratew 0.809 Average wage Unemployment rate = 6.1%ψ1 0.0309 Scale of hiring cost Recruiting costs of 3% of pay1/κ 0.0810 Search effort elasticity d log u/d log b|x = 0.5ξ 0.230 Pain from unemployment Leisure benefit of unemployment

Panel D. Parameters jointly calibrated to volatilities

ρ 0.9 Autocorrelation of shocks 0.9StDev(ηA) 0.724% TFP innovation StDev(ut) = 1.59%StDev(ηI) 0.244% Monetary policy innovation StDev(Y | ηA) = StDev(Y | ηI)StDev(ηG) 4.63% Government purchases innovation StDev(Gt/Yt) = 1.75%

ζ 1.68 Elasticity of wage w.r.t. x StDev(ht)/ StDev(1− ut) = 0.568

Panel E: Automatic stabilizers

b 0.783 UI replacement rate Rothstein and Valletta (2014)τ 0.151 Progressivity of tax system Heathcote et al. (2014)

Table 1: Calibrated parameter values and targets

28

in the US in 1984-2007. For the monetary policy rule, we use OLS estimates of equation (8). The

parameter υ determines the probability of not having a job in our model, which we set equal to

the sum of quarterly job separation rate, which we construct following Shimer (2012), and the

average unemployment rate. Finally, Guvenen and McKay (2016) estimate a version of the income

innovation process that we have specified in equation (3) using a mixture of normals as a flexible

parameterization of the distribution F (ε′; ·). Two of the mixtures shift with a measure of slack to

match the observed pro-cyclical skewness of earnings growth rates documented by Guvenen et al.

(2014), and we take the unemployment rate to be the measure of slack in our implementation.

Appendix E.1 provides additional details.

Panel C instead has parameters chosen jointly to target a set of moments. We target the

average unemployment rate between 1960 to 2014 and recruiting costs of 3 percent of quarterly

pay, consistent with Barron et al. (1997). The parameter κ controls the marginal disutility of

effort searching for a job, and we set it to target a micro-elasticity of unemployment with respect

to benefits of 0.5 as reported by Landais et al. (2015). Last in the panel is ξ, the non-pecuniary

costs of unemployment. In the model, the utility loss from unemployment is log(1/b)− h1+γ/(1 +

γ) + ξ, reflecting the loss in consumption, the gain in leisure, and other non-pecuniary costs of

unemployment. We set ξ = h1+γ/(1 + γ) in the steady state of our baseline calibration so that the

benefit of increased leisure in unemployment is dissipated by the non-pecuniary costs.

Panel D calibrates the three aggregate shocks in our model that perturb productivity, monetary

policy, and the rule for public expenditures. In each case we assume that the exogenous process is

an AR(1) in logs with common autocorrelation. We set the variances to match three targets: (i) the

standard deviation of the unemployment rate, (ii) the standard deviation of Gt/Yt, and (iii) equal

contributions of productivity and monetary shocks to the variance of aggregate output. Finally, we

set ζ = 1.68 to match the standard deviation of hours per worker relative to the standard deviation

of the employment-population ratio.

Finally, panel E has the baseline values for the automatic stabilizers. For τ we adopt the

estimate of 0.151 from Heathcote et al. (2014). For b we choose b1/(1−τ) = 0.75, consistent with

a 20-25% decline in household income during unemployment reported by Rothstein and Valletta

(2014). Note that this implies interpreting a household as having two workers only one of whom

becomes unemployed so b1/(1−τ) = 0.75 corresponds to a 50 percent replacement rate of one worker’s

29

Steady state unemployment rate Standard deviation of log output

Figure 2: Effects of changing b for τ = 0.26

income. Our focus is on the effect that aggregate shocks have on these choices, not on their levels.

As a check on the model’s performance, we computed the standard deviation of hours, output,

and inflation in the model, which are 0.74%, 1.67%, and 0.65%. The equivalent moments in the

US data 1960-2014 are 0.84%, 1.32%, and 0.55%.

6.2 Optimal unemployment insurance

Our first main quantitative result is that aggregate shocks increase the optimal b to 0.805 from

0.752 in the absence of aggregate shocks. Converting the values for b into pre-tax UI replacement

rates based on a two-earner household then we have an optimal replacement rate of 49 percent with

aggregate shocks as compared to 36 percent without.25

Figure 2 provides a first hint for why this effect is so quantitatively large. It shows the effects

of raising b on the steady state unemployment rate and on the volatility of log output. Raising

the generosity of unemployment benefits hurts the incentives for working, so unemployment rises

somewhat. However, it has a strong macroeconomic stabilizing effect.

Figure 3 provides a different way to look at this result. Equation (27), repeated her for conve-

25The conversion we apply here is to solve for x in (x/2 + 1/2)1−τ = b.

30

nience:

E0

∞∑t=0

βtWt =E0

∞∑t=0

βt[Ei log

(α1−τi,t

)− log

(Ei[α1−τi,t

])]+E0

∞∑t=0

βt [ut log b− log (1− ut + utb)]

+E0

∞∑t=0

βt

[log(Ct)− (1− ut)

h1+γt

1 + γ− υ q

1+κt

1 + κ+ χ log(Gt)− ξut

],

shows that welfare is the sum of three components: the welfare loss from income inequality that

results from skill risk, αi,t, the welfare loss from income inequality as a result of unemployment, and

the utility that derives from aggregates. Figure 3 plots each of these components in consumption

equivalent units and their sum, as we vary b, both with and without business cycles.26

As expected, higher b provides insurance by lowering idiosyncratic income risk (top-left panel)

and unemployment risk (top-right panel). More interesting, the presence of aggregate shocks has a

large effect on the first effect, but a negligible one on the second. The reason is that, by stabilizing

the business cycle, a higher b leads to less pre-tax income inequality. Idiosyncratic income risk

is persistent and it increases in recessions. The contribution of skill risk to welfare is strongly

non-linear in the state of the business cycle and mitigating the most severe recessions leads to a

considerable gain in welfare. Instead, the welfare loss from unemployment risk is close to linear in

the state of the business cycle so the presence of aggregate risk and the stabilizing effects of benefits

have little effect on this component of welfare.

The next two panels confirm that uninsured skill risk drives the results. The welfare coming

from aggregates, in the lower-left panel, falls with b because of the moral hazard effect on the

unemployment rate, but the presence of aggregate shocks has a modest effect on the slope of the

relation between welfare and b. The sum of social welfare, plotted in the lower-right panel of figure

3 mimics the patterns in the top-left panel. The optimal unemployment benefit is substantially

larger with aggregate shocks because it stabilizes the business cycle leading to welfare gains by

26The consumption equivalent units are relative to the social welfare function in the deterministic steady stateassociated with the policy parameters that are optimal without aggregate shocks. Suppose the additively decomposedsocial welfare function for a given policy is V A+V B +V C . Let V A+ V B + V C be the social welfare function of in thesteady state with the policy parameters that are optimal without aggregate shocks. If we rescaled the consumption ofall households by a factor 1+∆, the latter would become V A+ V B+ V C+log(1+∆)/(1−β). The top-left panel plots∆A = exp(1−β)(V A−V A)−1. The bottom-right panel plots ∆ = exp(1−β)(V A+V B+V C−V A−V B−V C)−1.Notice that 1 + ∆ = (1 + ∆A)(1 + ∆B)(1 + ∆C).

31

Figure 3: Decomposition of welfare in consumption equivalent units. τ = 0.26. With aggregateshocks (black) and without (gray). Vertical lines show the optimal b with and without aggregateshocks.

32

Steady state output (log deviation) Standard deviation of log output

Figure 4: Effects of changing τ for b = 0.80. The left panel shows log output as a deviation fromthe value at τ = 0.265. The right panel uses the same vertical axis as figure 2 for comparison.

reducing the fluctuations in uninsured risk.

6.3 Optimal income tax progressivity

Our second main quantitative result is that the optimal tax progressivity is almost unchanged at

0.264 relative to 0.265 without aggregate shocks.

The left panel of figure 4 shows the steady state level of output as a log deviation from the level

under the optimal policy without aggregate shocks. There is a strong negative effect of progressivity

on output that works through the disincentive effect on labor supply. At the optimal policy without

aggregate shocks, the welfare loss from reducing aggregate output is balanced by the welfare gain

from insuring skill risk. When we introduce aggregate shocks, the level of tax progressivity has

essentially no effect on the volatility of the business cycle as shown in the right-panel of figure 4

and therefore there is no stabilization benefit of raising progressivity.

6.4 Alternative specifications

Table 2 repeats our optimal policy calculation for alternative calibrations of the model in order to

confirm what is driving the conclusions and assess their robustness. Row (i) repeats our baseline

results for comparison.

Row (ii) shows that macroeconomic stabilization concerns have almost no effect on the optimal b

and τ when prices are flexible. This confirms the important role of aggregate demand and inefficient

33

b τ replacement rate

Aggregate shocks................. without with without with without with

(i) Baseline 0.752 0.805 0.265 0.264 0.36 0.49(ii) Flexible prices 0.752 0.751 0.265 0.265 0.36 0.36(iii) Aggressive monetary policy 0.752 0.768 0.265 0.265 0.36 0.40(iv) Acyclical purchases 0.752 0.797 0.265 0.257 0.36 0.47(v) Volatile business cycle 0.752 0.809 0.265 0.246 0.36 0.51(vi) Smaller demand shocks 0.752 0.814 0.265 0.256 0.36 0.52(vii) Larger demand shocks 0.752 0.792 0.265 0.243 0.36 0.47(viii) Positive wage elasticity 0.492 0.708 0.241 0.276 – 0.24

Table 2: Optimal policies under alternative specifications. Replacement rates are calculated basedon two-worker households as described in footnote 25.

business cycles.

Rows (iii) and (iv) investigate the interaction with other policies. Row (iii) increases the coeffi-

cient on inflation in the monetary policy rule to 2.50 from 1.66. With this more aggressive monetary

policy rule there is less of a need for fiscal policy to manage aggregate demand. Therefore, stabi-

lization plays a smaller role in the design of the optimal social insurance system than in the baseline

calibration. Another way to understand these results is in terms of the slope of the AD curve in

figure 1. Flexible prices or aggressive monetary policy make the real interest rate responds strongly

to changes in slack. From Lemma 3 then, the AD curve will be flatter, and so the elasticities of

slack with respect to the social programs is small leading to a small automatic-stabilizer role.

Row (iv) changes instead the policy rule for government purchases. Our baseline specification,

following the Samuelson rule, makes government purchases pro-cyclical. Row (iv) considers instead

acyclical government purchases, as we observe in the data, by using instead the rule: Gt = GηGt .

Because the pro-cyclical rule amplified the business cycle it left a larger role for the automatic

stabilizers, but the effect is quantitatively minor.

Rows (v), (vi), and (vii) show that the calibration of the aggregate shocks is not crucial to our

results. Row (v) raises the standard deviations of the productivity and monetary policy shocks

by 25%. As one might expect, when the aggregate shocks are more volatile, stabilization plays a

larger role in the choice of b and τ . In our baseline calibration, the productivity and monetary

policy shocks contribute equally to the variance of log output. In row (vi) we consider the case

where productivity shocks account for 75% of the variance and monetary policy shocks account for

34

25% and in row (vii) we consider the reverse case where productivity shocks account for 25%. As

the ratio of the variance of monetary relative to productivity shocks rises, the automatic stabilizer

role of social insurance programs diminishes but only slightly. To understand these findings, again

consider Lemma 3. The precautionary savings motive rises as the unemployment rate increases,

which leads to a steeper AD curve. This implies that any shifts in the AS or AD curves will

have larger effects on slack. This time-varying precautionary motive is an important part of our

quantitative analysis and unemployment insurance has a strong stabilizing effect on the economy

by interfering with this self-reinforcing mechanism. As unemployment insurance flattens the AD

curve, it leads to smaller fluctuations in slack for any type of shock.

Finally, row (viii) considers an alternative specification of the wage mechanism in which w

responds to increases in the unemployment benefit. Our baseline specification assumes that the

steady state wage is essentially unchanged across policy regimes.27 This specification is consistent

with empirical work that fails to find an effect of unemployment benefits on wages.28 However, some

authors have found large effects of unemployment benefits on the equilibrium unemployment rate

possibly reflecting general equilibrium effects operating through wages (Hagedorn et al., 2013). We

explore how our analysis is affected when the steady state wage is increasing in the unemployment

benefit.

In this specification we assume that w has an elasticity of 0.1 with respect to b. As one would

expect, with a positive wage elasticity the steady state unemployment rate increases with the

benefits more strongly than before. Higher benefits raise wages, which lowers the incentives for

firms to hire workers leading to a lower job-finding rate. Moreover, the incentives for searching for

a job fall both due to the decline in the job-finding rate and because of the usual moral hazard

effect. With a positive wage elasticity, the greater sensitivity of the steady state unemployment

rate to the unemployment benefit leads to a much lower optimal benefit in the absence of aggregate

shocks: 0.49 as opposed to 0.71. However, when there are aggregate shocks, a very low benefit leads

to strong de-stabilizing dynamics because of the precautionary savings motive. Such low levels of

the benefit lead to very costly fluctuations, so the automatic-stabilizer nature of unemployment

benefits is much stronger. As a result, in this case we find that aggregate shocks lead to a large

27It is not exactly constant due to changes in the term 1− Jt/Yt although in practice this term is small.28Card et al. (2007), Van Ours and Vodopivec (2008), Lalive (2007), and Johnston and Mas (2015) find no evidence

that UI generosity affects earnings upon re-employment.

35

change in the optimal benefit relative to what is optimal in steady state. A positive wage elasticity

with respect to benefits leads to lower replacement rates overall, but a bigger effect of the business

cycle on the optimal policy.

7 Conclusion

Policy debates take as given that there are stabilizing benefits of unemployment insurance and

income tax progressivity, but there are few systematic studies of what factors drive these bene-

fits and how large they are. The study of these social programs rarely takes into account this

macroeconomic stabilization role, instead treating it a fortuitous side benefits.

This paper tried to remedy this situation. Our theoretical study provided a theoretical charac-

terization of what is an automatic stabilizer. In general terms, an automatic stabilizer is a fixed

policy for which there is a positive covariance between the effect of slack on welfare, and the effect

of the policy on slack. If a policy tool has this property of stimulating the economy more in re-

cessions when slack is inefficiently high, then its role in stabilizing the economy calls for expanding

the use of the policy beyond what would be appropriate in a stationary environment. We showed

what factors drive this covariance, through which economic channels they operate, and what makes

them larger or smaller. Overall, we found that the role of social insurance programs as automatic

stabilizers affects their optimal design and, in the case of unemployment insurance, it can lead to

substantial differences in the generosity of the system.

Our focus on the automatic stabilizing nature of existing social programs led us to take a Ramsey

approach to the ex ante design of fiscal policy. A different question is whether macroeconomic

stabilization concerns interfere with the incentive problems that give rise to these programs, as in

the Mirrleesian literature. Another question is how these fiscal policy programs can adjust to the

state of the business cycle, taking into account measurement difficulties, time inconsistency, and

the many other issues that the literature on monetary policy has faced. There is already some

progress on these two topics, and hopefully our analysis will provide some insights to its further

development.

36

References

Aiyagari, S. R. and McGrattan, E. R. (1998). The optimum quantity of debt. Journal of Monetary

Economics, 42(3):447–469.

Auclert, A. (2016). Monetary policy and the redistribution channel. Princeton University

manuscript.

Auerbach, A. J. and Feenberg, D. (2000). The significance of federal taxes as automatic stabilizers.

Journal of Economic Perspectives, 14(3):37–56.

Baily, M. N. (1978). Some aspects of optimal unemployment insurance. Journal of Public Eco-

nomics, 10(3):379–402.

Barron, J. M., Berger, M. C., and Black, D. A. (1997). On-the-job training. W.E. Upjohn Institute

for Employment Research, Kalamazoo, MI.

Basu, S. and Fernald, J. G. (1997). Returns to scale in us production: Estimates and implications.

Journal of Political Economy, 105(2):249–283.

Benabou, R. (2002). Tax and education policy in a heterogeneous agent economy: What levels of

redistribution maximize growth and efficiency? Econometrica, 70(2):481–517.

Bhandari, A., Evans, D., Golosov, M., and Sargent, T. J. (2013). Taxes, debts, and redistributions

with aggregate shocks. NBER working paper 19470.

Blanchard, O., Dell’Ariccia, G., and Mauro, P. (2010). Rethinking macroeconomic policy. Journal

of Money, Credit and Banking, 42(s1):199–215.

Blanchard, O. and Galı, J. (2010). Labor markets and monetary policy: A New Keynesian model

with unemployment. American Economic Journal: Macroeconomics, 2(2):1–30.

Calvo, G. A. (1983). Staggered prices in a utility-maximizing framework. Journal of Monetary

Economics, 12(3):383 – 398.

Card, D., Chetty, R., and Weber, A. (2007). Cash-on-hand and competing models of intertemporal

behavior: New evidence from the labor market. Quarterly Journal of Economics, 122(4):1511–

1560.

37

Chari, V. V., Kehoe, P. J., and McGrattan, E. R. (2007). Business cycle accounting. Econometrica,

75(3):781–836.

Chetty, R. (2006). A general formula for the optimal level of social insurance. Journal of Public

Economics, 90(10):1879–1901.

Chetty, R. (2012). Bounds on elasticities with optimization frictions: A synthesis of micro and

macro evidence on labor supply. Econometrica, 80(3):969–1018.

Conesa, J. C. and Krueger, D. (2006). On the optimal progressivity of the income tax code. Journal

of Monetary Economics, 53(7):1425–1450.

De Santis, M. (2007). Individual consumption risk and the welfare cost of business cycles. American

Economic Review, 97(4):1488 – 1506.

Den Haan, W. J., Rendahl, P., and Riegler, M. (2015). Unemployment (fears) and deflationary

spirals. CEPR Discussion Paper 10814.

DiMaggio, M. and Kermani, A. (2016). The importance of unemployment insurance as an automatic

stabilizer. Columbia Business School Research Paper 15-31.

Gali, J. (2011). Unemployment Fluctuations and Stabilization Policies: A New Keynesian Perspec-

tive. MIT Press, Cambridge, USA.

Galı, J., Gertler, M., and Lopez-Salido, J. D. (2007). Markups, gaps, and the welfare costs of

business fluctuations. The Review of Economics and Statistics, 89(1):44–59.

Golosov, M., Troshkin, M., and Tsyvinski, A. (2016). Redistribution and social insurance. American

Economic Review, 106(2):359–386.

Guerrieri, V. and Lorenzoni, G. (2011). Credit crises, precautionary savings, and the liquidity trap.

NBER Working Paper 17583.

Guvenen, F., Karahan, F., Ozkan, S., and Song, J. (2015). What do data on millions of U.S.

workers reveal about life-cycle earnings risk? University of Minnesota working paper.

Guvenen, F. and McKay, A. (2016). A parsimonious income process for business cycle analysis.

Boston University Working Paper.

38

Guvenen, F., Ozkan, S., and Song, J. (2014). The nature of countercyclical income risk. Journal

of Political Economy, 122(3):pp. 621–660.

Hagedorn, M., Karahan, F., Manovskii, I., and Mitman, K. (2013). Unemployment benefits and

unemployment in the great recession: The role of macro effects. NBER Working Paper 19499.

Hall, R. E. (2005). Employment fluctuations with equilibrium wage stickiness. American Economic

Review, 95(1):50–65.

Heathcote, J. and Perri, F. (2015). Wealth and volatility. NBER Working Paper 20944.

Heathcote, J., Storesletten, K., and Violante, G. L. (2014). Optimal Tax Progressivity: An Ana-

lytical Framework. NBER working paper 19899.

Hosios, A. J. (1990). On the efficiency of matching and related models of search and unemployment.

Review of Economic Studies, 57(2):279–298.

IMF (2015). Can fiscal policy stabilize output? In Fiscal Monitor, chapter 2, pages 21–49.

International Monetary Fund.

Johnston, A. C. and Mas, A. (2015). Potential unemployment insurance duration and labor supply:

the individual and market-level response to a benefit cut. Princeton University working paper.

Jung, P. and Kuester, K. (2015). Optimal labor-market policy in recessions. American Economic

Journal: Macroeconomics, 7(2):124–156.

Kaplan, G. W., Moll, B., and Violante, G. (2016). Monetary policy according to HANK. Princeton

University manuscript.

Kekre, R. (2015). Unemployment insurance in macroeconomic stabilization. Harvard University

manuscript.

Klenow, P. J. and Malin, B. A. (2010). Microeconomic evidence on price-setting. In Friedman,

B. and Woodford, M., editors, Handbook of Monetary Economics, volume 3A, pages 231–284.

Elsevier: North Holland.

Krebs, T. (2003). Growth and welfare effects of business cycles in economies with idiosyncratic

human capital risk. Review of Economic Dynamics, 6(4):846–868.

39

Krebs, T. (2007). Job displacement risk and the cost of business cycles. American Economic

Review, 97(3):664–686.

Krueger, D. and Ludwig, A. (2013). Optimal progressive labor income taxation and education

subsidies when education decisions and intergenerational transfers are endogenous. American

Economic Review, 103(3):496–501.

Krusell, P., Mukoyama, T., and Smith, Jr., A. A. (2011). Asset prices in a Huggett economy.

Journal of Economic Theory, 146(3):812–844.

Lalive, R. (2007). Unemployment benefits, unemployment duration, and post-unemployment jobs:

A regression discontinuity approach. American Economic Review, 97(2):108–112.

Landais, C., Michaillat, P., and Saez, E. (2015). A macroeconomic approach to optimal unemploy-

ment insurance: Theory. LSE manuscript.

Maliar, L. and Maliar, S. (2015). Merging simulation and projection approaches to solve high-

dimensional problems with an application to a new Keynesian model. Quantitative Economics,

6(1):1–47.

Mankiw, N. G. and Reis, R. (2010). Imperfect information and aggregate supply. In Friedman,

B. and Woodford, M., editors, Handbook of Monetary Economics, volume 3A, chapter 5, pages

183–230. Elsevier: North Holland.

McKay, A., Nakamura, E., and Steinsson, J. (2016). The power of forward guidance revisited.

American Economic Review, forthcoming.

McKay, A. and Reis, R. (2016). The role of automatic stabilizers in the U.S. business cycle.

Econometrica, 84(1):141–194.

Mirrlees, J. A. (1971). An exploration in the theory of optimum income taxation. Review of

Economics Studies, 38(2):175–208.

Mitman, K. and Rabinovich, S. (2011). Pro-cyclical unemployment benefits? optimal policy in an

equilibrium business cycle model. PIER working paper 11-010.

40

Musgrave, R. A. and Miller, M. H. (1948). Built-in flexibility. American Economic Review,

38(1):122–128.

Oh, H. and Reis, R. (2012). Targeted transfers and the fiscal response to the great recession.

Journal of Monetary Economics, 59(Supplement):S50–S64.

Perotti, R. (2005). Estimating the effects of fiscal policy in OECD countries. Federal Reserve Bank

of San Francisco, Conference Proceedings.

Petrongolo, B. and Pissarides, C. A. (2001). Looking into the black box: A survey of the matching

function. Journal of Economic Literature, 39:390–431.

Ravn, M. and Sterk, V. (2013). Job uncertainty and deep recessions. University College London

working paper.

Rothstein, J. and Valletta, R. G. (2014). Scraping by: Income and program participation after the

loss of extended unemployment benefits. IZA Discussion Paper No. 8022.

Samuelson, P. A. (1954). The pure theory of public expenditure. The Review of Economics and

Statistics, 36(4):387–389.

Shimer, R. (2012). Reassessing the ins and outs of unemployment. Review of Economic Dynamics,

15(2):127 – 148.

Storesletten, K., Telmer, C. I., and Yaron, A. (2001). The welfare cost of business cycles revisited:

Finite lives and cyclical variation in idiosyncratic risk. European Economic Review, 45(7):1311–

1339.

Storesletten, K., Telmer, C. I., and Yaron, A. (2004). Cyclical dynamics in idiosyncratic labor

market risk. Journal of Political Economy, 112(3):695–717.

Van Ours, J. C. and Vodopivec, M. (2008). Does reducing unemployment insurance generosity

reduce job match quality? Journal of Public Economics, 92(3):684–695.

Varian, H. R. (1980). Redistributive taxation as social insurance. Journal of Public Economics,

14(1):49–68.

41

Werning, I. (2007). Optimal fiscal policy with redistribution. Quarterly Journal of Economics,

122(3):925–967.

Werning, I. (2015). Incomplete markets and aggregate demand. NBER Working Paper 21448.

Woodford, M. (2010). Optimal monetary stabilization policy. In Friedman, B. M. and Woodford,

M., editors, Handbook of Monetary Economics, volume 3, chapter 14, pages 723 – 828. Elsevier:

North Holland.

42

Appendix

This appendix contains five sections, which: describe the role of the wage rule introduced in

section 2; provide auxiliary steps to some results stated in section 3; prove the propositions in

sections 4; derive the AD-AS apparatus and prove the propositions in section 5; and describe the

methods used to solve the model in section 6.

A The role of the wage rule in our analysis

We assumed that the wage was determined by equation (6). A more general specification uses a

generic wage function:

wt = w(ηt, xt, b, τ),

that maps the aggregate shocks, slack (or labor market tightness), and policy parameters into a

wage. Nash bargaining is an example of a wage mechanism that would take this form.

In this more general case, starting from (44) we can write ht as:

ht =

(1− τ)

[ηAtS(xt)

(1− Jt

Yt

)]−1w(ηAt , xt, b, τ)

1/(1+γ)

where S(xt) is the level of price dispersion associated with that level of slack. Using the generic

wage rule and equations (4), (5), (21), (25), (42), and (44), we can write Jt/Yt as a function of

(ηt, xt, b, τ) to yield:

ht = (1− τ)H(ηt, xt, b, τ)1/(1+γ) ,

where Ht ≡[

ηtS(xt)

(1− Jt

Yt

)]−1w(ηt, xt, b, τ). The purpose of assuming the wage rule in equation

(6) in the paper is that Ht simplifies to xζt .

However, not making this simplification, and so carrying the new H(.) term along only adds

to our results a few additional terms. To start, in addition to those terms that appear in equation

43

(28) (Proposition 1), we would now need to add:

E0

∞∑t=0

βt(1− ut)[AtCt− hγt

]dhtdHt

∂Ht

db

∣∣∣∣x

.

Similarly, we would need to modify equation (29) (Proposition 2) to include a term:

E0

∞∑t=0

βt(1− ut)[AtCt− hγt

]dhtdHt

∂Ht

dτ

∣∣∣∣x

.

Both of these terms relate to the adjustment of hours on the intensive margin. Intuitively, the

wage plays two roles in the economy: it determines the marginal cost of the firms and therefore

the incentives for hiring and it determines the incentives for labor supply on the intensive margin.

The first effect is captured in our analysis through dxt/db. Using our wage rule, the second effect

is captured through dhtdxt

dxtdb . With a more general wage rule brings in these new effects of policy on

intensive margin labor supply. While they result in a different value for dxt/db and dxt/dτ , they

are not qualitatively different considerations than the ones we focus on already.

B Additional steps for deriving the results in section 3

B.1 The value of employment

In equilibrium, ai,t = 0, and search effort is determined by comparing the value of working and not

working according to equation (16). This section of the appendix derives the two key steps that

make this difference independent of the household’s type, so that all households choose the same

search effort.

Lemma 4. Suppose the household’s value function has the form

V (α, n,S) = V α(α,S) + V n(a, n,S)

for some functions V α and V n. The choice of search effort is then the same for all searching

households regardless of α.

44

Proof: The household’s search problem is

V s(α,S) = maxq

MqV (α, 1,S) + (1−Mq)V (α, 0,S)− q1+κ

1 + κ

.

Substitute for the value functions to arrive at

V s(α,S) = maxq

Mq [V α(α,S) + V n(a, 1,S)] + (1−Mq) [V α(α,S) + V n(a, 0,S)]− q1+κ

1 + κ

V s(α,S) = V α(α,S) + max

q

MqV n(1,S) + (1−Mq)V n(0,S)− q1+κ

1 + κ

, (36)

where we have brought V α(α,S) outside the max operator as it appears in an additively separable

manner. As there is no α inside the max operator, the optimal q is independent of α.

Lemma 5. The household’s value function can be written as

V (α, n,S) =1− τ1− β

log(α) + n

[log

1

b− h1+γ

1 + γ+ ξ

]+ V (S) (37)

for some function V .

Proof. Suppose that the value function is of the form given in (37). We will establish that the

Bellman equation maps functions in this class into itself, which implies that the fixed point of the

Bellman equation is in this class by the contraction mapping theorem. V s will then be

V s(α,S) =1− τ1− β

log(α) +Mq∗[log

1

b− h1+γ

1 + γ+ ξ

]− q∗1+κ

1 + κ+ V (S)

and the choice of q∗ is independent of α by Lemma 4. Regardless of employment status, the

continuation value is

(1− υ)V (α′, 1,S ′) + υV s(α′,S ′)

= (1− υ)

[1− τ1− β

log(α′) +

[log

1

b− h′1+γ

1 + γ+ ξ

]+ V (S ′)

]

+ υ

[1− τ1− β

log(α′) +M ′q∗′[

log1

b− h′1+γ

1 + γ+ ξ

]− q∗′

1+κ

1 + κ+ V (S ′)

]=

1− τ1− β

log(α′) + g(S ′),

45

where

g(S) ≡ (1− υ + υM ′q∗′)

[log

1

b− h′1+γ

1 + γ

]− (1− υ)

q∗′1+κ

1 + κ+ V (S ′).

Turning to the Bellman equation, we have

V (α, n,S) = log[λ(n+ (1− n)b) (α(wh+ d))1−τ

]− n h

1+γ

1 + γ− (1− n)ξ + βE

[1− τ1− β

log(α′) + g(S ′)]

=1− τ1− β

log(α) + n

[log

1

b− h1+γ

1 + γ+ ξ

]+ log

[λ (wh+ d)1−τ

]+ log b− ξ + β

1− τ1− β

E[log(ε′)

]+ βE

[g(S ′)

].

Finally, V (S) is given by the second row of the expression above.

B.2 Proof of lemma 2

First, the Euler equation for a household is

1

ci,t≥ βRtE

[1

ci,t+1

].

as usual. At the same time, as we showed in the text, the consumption of an individual is

ci,t = λtα1−τi,t (wtht + dt)

1−τ (ni,t + (1− ni,t)b). (38)

Replacing individual consumption into the Euler equation, and rewriting gives:

1

λt (wtht + dt)1−τ ≥ βRtEt

(1− ut+1) + ut+1b

−1

λt+1 (wt+1ht+1 + dt+1)1−τ

Et

[α1−τi,t

α1−τi,t+1

][ni,t + (1− ni,t)b] . (39)

Notice that E[α1−τi,t

α1−τi,t+1

]= E

[ετ−1i,t+1

]is common across households. This Euler equation only dif-

fers across households due to the final term involving ni,t. Assuming the unemployment benefit

replacement rate is less than one, this term will be larger for employed than unemployed so in

equilibrium all unemployed will be constrained and the Euler equation will hold with equality for

all employed.29

29Here we follow Krusell et al. (2011), Ravn and Sterk (2013), and Werning (2015) in assuming that the Eulerequation of the employed/high-income household holds with equality. This household is up against its constrainta′ = 0 so there could be other equilibria in which the Euler equation does not hold with equality. The equilibrium

46

Summing equation (38) across households gives:

Ct = λt (wtht + dt)1−τ (1− ut + utb)Ei

[α1−τi,t

]. (40)

Solve (40) for λt (wtht + dt)1−τ and substitute it into (38) to arrive at:

ci,t = α1−τi,t (ni,t + (1− ni,t)b)

Ct

(1− ut + utb)Ei[α1−τi,t

] . (41)

Equations (17) and (22) follow from (41) and the definition of ct.

To derive equations (18) and (19), solve (40) for λt (wtht + dt)1−τ and substitute it into (39)

holding with equality and ni,t = 1 to arrive at:

(1− ut + utb)Ei[α1−τi,t

]Ct

= βRtEt

[(1− ut+1) + ut+1b−1] (1− ut+1 + ut+1b)Ei

[α1−τi,t+1

]Ct+1

E[ετ−1i,t+1

]

and substitute in ct using (17):

1

ct= βRtEt

[(1− ut+1) + ut+1b

−1] 1

ct+1

E[ετ−1i,t+1

].

Rearranging gives eqations (18) and (19) as E[ετ−1i,t+1

]is known at date t.

B.3 Optimal hours worked and search effort

We start by deriving equation (20). Using labor market clearing and the definition of At we can

write the aggregate production function as

Yt = Atht(1− ut). (42)

we focus on is the limit of the unique equilibrium as the borrowing limit approaches zero from below. See Krusellet al. (2011) for further discussion of this point.

47

Output net of hiring costs is paid to employed workers in the form of wage and dividend payments.

As the average αi,t is equal to one we have

Yt − Jt = (wtht + dt )( 1− ut) . (43)

Multiply both sides of (42) by (Yt − Jt)/Yt and substitute for Yt − Jt using (43) to give:

wtht + dt = AthtYt − JtYt

.

Substitute this for wtht + dt in equation (15) to arrive at

hγt =(1− τ)wt

AthtYt−JtYt

. (44)

Finally, use the wage rule in equation (6) to arrive at (20).

We turn next to derive optimal search effort in equation (21). By Lemma 5, the results of

Lemma 4 can be applied. Proceed from equation (36) using the functional form for the value

function in Lemma 5. This leads to

maxq

Mq

[log

1

b− h1+γ

1 + γ+ ξ

]− q1+κ

1 + κ

.

The first order condition yields equation (21).

B.4 Equilibrium definition

We first state some addition equilibrium conditions and then state a definition of an equilibrium.

Here we focus on the Calvo-pricing version of the model for concreteness.

The aggregate resource constraint is:

Yt − Jt = Ct +Gt. (45)

The Fisher equation is:

Rt = It/Et [πt+1] . (46)

48

The price-setting first order condition leads to:

p∗tpt

=Et∑∞

s=tR−1t,s (1− θ)s−t

(ptps

)µ/(1−µ)Ysµ

(wshs + ψ1M

ψ2s

)/(Ashs)

Et∑∞

s=tR−1t,s (1− θ)s−t

(ptps

)1/(1−µ)Ys

. (47)

The link between ct and Ct depends on Ei[α1−τi,t

]. This evolves according to:

Ei[α1−τi,t

]= (1− δ)Ei

[α1−τi,t−1

]Ei[ε1−τi,t

]+ δ.. (48)

An equilibrium of the economy can be calculated from a system equations in 17 variables and

three exogenous processes. The variables are

Ct, ct, ut,Ei[α1−τi,t

], Qt, Rt, It, πt, Yt, Gt, ht, wt, St,

p∗tpt, Jt, qt,Mt.

And the equations are: (4), (5), (6), (7), (8), (9), (17) (18), (19), (20), (21), (24), (42), (45), (46),

(47), and (48). The exogenous processes are ηAt , ηGt , and ηIt .

C Proofs for section 4

C.1 Skill dispersion with log normality

Lemma 6. Under no mortality, δ = 0, and a log normal income process:

E0

∞∑t=0

βt Ei[log(α1−τi,t

)− log

(Ei[α1−τi,t

])]= log

(α1−τi,0

)−log

(Ei[α1−τi,0

])−E0

∞∑t=0

βtβ

1− β(1−τ)2

σ2(xt)

2

Proof: When there is no mortality, δ = 0, we can compute the cumulative welfare effect of a change

in F (εi,t+1, xt) including the effects on current and future skill dispersion. In particular

Ei log(α1−τi,t

)= Ei log

(α1−τi,t−1ε

1−τi,t

)= Ei log

(α1−τi,0 ε1−τi,1 · · · ε

1−τi,t

)= Ei log

(α1−τi,0

)+ Ei log

(ε1−τi,1

)+ · · ·+ Ei log

(ε1−τi,t

).

49

Similarly

log(Ei[α1−τi,t

])= log

(Ei[α1−τi,t−1

]Ei[ε1−τi,t

])= log

(Ei[α1−τi,0

]Ei[α1−τi,1

]· · ·Ei

[ε1−τi,t

])= log

(Ei[α1−τi,0

])+ log

(Ei[α1−τi,1

])+ · · ·+ log

(Ei[ε1−τi,t

])Notice that in this no-mortality case, the date-t loss from skill dispersion can be written as:

Ei log(α1−τi,t

)− log

(Ei[α1−τi,t

])= Ei log

(α1−τi,0

)− log

(Ei[α1−τi,0

])+

t∑s=1

[Ei log

(ε1−τi,s

)− log

(Ei[ε1−τi,s

])].

The level of slack at date t, xt, determines the contribution Ei log(ε1−τi,t+1

)− log

(Ei[ε1−τi,t+1

])that

will appear in Ws for all future dates s > t. Using the log-normal income process, we have

Ei log(ε1−τi,t+1

)− log

(Ei[ε1−τi,t+1

])= −(1− τ)

σ2(xt)

2+ (1− τ)

σ2(xt)

2− (1− τ)2

σ2(xt)

2= −(1− τ)2

σ2(xt)

2.

So the effect of xt on∑∞

t=0 βtWt is given by

− β

1− β(1− τ)2

σ2(xt)

2, (49)

proving the result.

C.2 Proof of Proposition 1

For this proof, in addition to the social welfare function, (27), the relevant equations of the model

are (4), (42), (5), (45), (21), and (20).

The first-order condition of the social welfare function with respect to b is

E0

∞∑t=0

βtutb− ut

1− ut + utb+dWt

dut

∂ut∂b

∣∣∣∣x

− υqκt∂qt∂b

∣∣∣∣x

+dWt

dxt

dxtdb

= 0. (50)

50

The first two terms in (50) can be expressed as

utb− ut

1− ut + utb= ut

(1

b− 1 + 1− 1

1− ut + utb

)= ut

(1

b− 1

)(1− utb

1− ut + utb

)

and note that

∂ log (bct)

∂ log b

∣∣∣∣x,q

=∂

∂ log blog

bCt

Ei[α1−τi,t

](1− ut + utb)

(51)

= 1− utb

1− ut + utb

where the partial derivative on the right hand side of (51) is with respect to b alone.30 So, we have:

utb− ut

1− ut + utb= ut

(1

b− 1

)∂ log (bct)

∂ log b

∣∣∣∣x,q

(52)

For the third and fourth terms in (50), start by noting that:

dWt

dut= log b+

1− b1− ut + utb

− AthtCt

+ψ1M

ψ2t

Ct+h1+γt

1 + γ− ξ

=

(log b+

h1+γt

1 + γ− ξ

)+

1

ct

dctdut

(53)

where

1

ct

dctdut

=1− b

1− ut + utb− Atht

Ct+ψ1M

ψ2t

Ct.

Then, it follows that:

dWt

dut

∂ut∂b

∣∣∣∣x

− υqκt∂qt∂b

∣∣∣∣x

=

[(log b+

h1+γt

1 + γ− ξ

)+

1

ct

dctdut

]∂ut∂b

∣∣∣∣x

− υqκt∂qt∂b

∣∣∣∣x

.

30As Ei[α1−τi,t

]is an endogenous state that depends on the history of x, we are taking the partial derivative holding

fixed this history.

51

Using equation (21), this becomes

dWt

dut

∂ut∂b

∣∣∣∣x

− υqκt∂qt∂b

∣∣∣∣x

=

[(log b+

h1+γt

1 + γ− ξ

)+

1

ct

dctdut

]∂ut∂b

∣∣∣∣x

+

(− log b− h1+γt

1 + γ+ ξ

)dutdqt

∂qt∂b

∣∣∣∣x

=1

ct

dctdut

∂ut∂b

∣∣∣∣x

=d log ctd log ut

∂ log ut∂b

∣∣∣∣x

. (54)

Substituting (52) and (54) into (50) yields the result.

C.3 Proof of Proposition 2

We proceed as in the proof of Proposition 1. The first-order condition of the social welfare function

with respect to τ is:

E0

∞∑t=0

βt

β

1− β(1− τ)σ2(xt) +

dWt

dht

∂ht∂τ

∣∣∣∣x

+dWt

dxt

dxtdτ

= 0. (55)

From (27), (42), and (45) we have

dWt

dht=dWt

dut

dutdqt

dqtdht

+At(1− ut)

Ct− (1− ut)hγt − υqκt

dqtdht

.

From (21) we have

∂ht∂τ

∣∣∣∣x

= − ht(1− τ)(1 + γ)

combining these and using (21), (53), and dut/dqt = −υMt we arrive at

dWt

dht

∂ht∂τ

∣∣∣∣x

=dWt

dut

dutdqt

dqtdht

∂ht∂τ

∣∣∣∣x

+At(1− ut)

Ct

∂ht∂τ

∣∣∣∣x

− (1− ut)hγt∂ht∂τ

∣∣∣∣x

− υqκtdqtdht

∂ht∂τ

∣∣∣∣x

=1

ct

dctdut

dutdqt

dqtdht

∂ht∂τ

∣∣∣∣x

−[Atht(1− ut)

Ct− (1− ut)h1+γt

]1

(1− τ)(1 + γ)

=d log ctd log ut

∂ log ut∂τ

∣∣∣∣x

−[AtCt− hγt

]ht(1− ut)

(1− τ)(1 + γ).

Substituting this into (55) yields the desired result.

52

D Proofs for section 5

D.1 Proof of Proposition 3

Proceeding as in the proof of Proposition 1 we have

dWt

dxt=dWt

dut

dutdxt− υqκt

dqtdxt

+

[At(1− ut)

Ct− (1− ut)hγt

]dhtdxt− 1

Ct

∂Jt∂xt

∣∣∣∣u

(56)

− β

1− β(1− τ)2

d

dxt

σ2(xt)

2− YtCtSt

dStdxt

where dWt/dut is given by (53). We rearrange (56) to arrive at the desired result. First, note that:

dutdxt

= −υqtdMt

dxt− υMt

dqtdxt

. (57)

Using this, (21), and (53), equation (56) then becomes

dWt

dxt=

(−

(− log b− h1+γt

1 + γ+ ξ

)+

1

ct

dctdut

)(−υqt

dMt

dxt− υMt

dqtdxt

)− υMt

(− log b− h1+γt

1 + γ+ ξ

)dqtdxt

+

[At(1− ut)

Ct− (1− ut)hγt

]dhtdxt− 1

Ct

∂Jt∂xt

∣∣∣∣u

− β

1− β(1− τ)2

d

dxt

σ2(xt)

2− YtCtSt

dStdxt

= υqt

(− log b− h1+γt

1 + γ+ ξ

)dMt

dxt− YtCtSt

dStdxt

+1

ct

dctdut

dutdxt

+

[At(1− ut)

Ct− (1− ut)hγt

]dhtdxt− 1

Ct

∂Jt∂xt

∣∣∣∣u

− β

1− β(1− τ)2

d

dxt

σ2(xt)

2.

This can be rearranged to the desired result by making use of

1

ct

dctdut

=1

Ct

dCtdut

+1− b

1− ut + utb,

and

−υqtdMt

dxt=∂ut∂xt

∣∣∣∣q

.

53

D.2 Deriving the special case in section 5.2.3

We can normalize Et−1pt = 1. A fraction θ of firms set the price p∗t and the remaining set the price

Et−1p∗t = 1. From the definition of the price index we can solve for

(p∗tpt

)µ/(1−µ)=

(1− (1− θ)p−1/(1−µ)t

θ

)µ.

Substituting into (25) we arrive at

St = (1− θ)pµ/(µ−1)t + θ1−µ(

1− (1− θ)p1/(µ−1)t

)µ.

This makes clear that St is a function of pt. Differentiating this function we arrive at

S′(1) = (1− θ) µ

µ− 1

[p1/(µ−1)t − θ1−µ

(1− (1− θ)p1/(µ−1)t

)µ−1p(2−µ)/(µ−1)t

]= 0

S′′(1) =1− θθ

µ

µ− 1.

Next, rewrite the welfare function as

Wt = log (Ct)− (1− ut)h1+γt

1 + γ+ χ log (Gt) + t.i.p.,

Ct = Yt − Jt −Gt,

Yt =ηAtS (pt)

ht(1− ut)

Gt = χCt.

Observe that ut and Jt are exogenous in this case as Mt and qt are exogenous. Use the production

function to rewrite ht in terms of Yt

ht =YtS (pt)

ηAt (1− ut).

Now rewrite Wt in terms of Yt and pt

Wt = (1 + χ) log

(Yt − Jt1 + χ

)− 1− ut

1 + γ

(YtS (pt)

ηAt (1− ut)

)1+γ

.

With flexible prices we have p = 1 and S = 1. The flexible-price, socially-efficient level of

54

output, Y ∗, satisfiesηA

Y ∗t − Jt −Gt=

(Y ∗t

ηAt (1− ut)

)γ.

Treating Wt as a function of Yt and pt, we take a second-order Taylor approximation around the

point (Y ∗t , 1). The first-derivatives are

WY (Y ∗t , 1) =1

Yt − Jt −Gt−(

YtS (pt)

ηAt (1− ut)

)γ S (pt)

ηAt

Wp(Y∗t , 1) = −(1− ut)

(YtS (pt)

ηAt (1− ut)

)γ Yt

ηAt (1− ut)S′ (pt)

and both are zero, the former because Y ∗ is optimal and the latter because S′(1) = 0. The second

derivatives are

WY Y (Y ∗t , 1) = − 1

(C∗)2− γ(Y ∗t )γ−1

(S (pt)

ηAt

)1+γ ( 1

1− ut

)γ,

WY p(Y∗t , 1) = −

(Y ∗t

1− ut

)γ ( 1

ηAt

)1+γ

(1 + γ)S(pt)γS′(pt),

Wpp(Y∗t , 1) = −(1− ut)

(Y ∗t

ηAt (1− ut)

)1+γ [γS (pt)

γ−1 S′ (pt) + S (pt)γ S′′(pt)

].

Using S(1) = 1, S′(1) = 0, the expression for S′′(1) above, and the optimality condition for Y ∗t we

arrive at

WY Y (Y ∗t , 1) = − 1

(C∗)2

[1 + γ

C∗tY ∗t

],

WY p(Y∗t , 1) = 0,

Wpp(Y∗t , 1) = −Y

∗t

C∗t

1− θθ

µ

µ− 1.

By Taylor’s theorem we can write

W (Yt, pt) ≈1

2WY Y (Y ∗t , 1)(Yt − Y ∗t )2 +

1

2Wpp(Y

∗t , 1)(pt − 1)2

Observe that Y and p are functions of x. So when we differentiate with respect to x we arrive at

dWt

dxt≈WY Y (Y ∗t , 1)(Yt − Y ∗t )

dYtdxt

+Wpp(Y∗t , 1)(pt − 1)

dptdxt

.

55

Substituting for WY Y and Wpp and rearranging yields the result.

D.3 The AD-AS apparatus in section 5.3

The first step in the analysis is to express the equilibrium in date 0 in terms of as few variables

as possible, namely x0. First, it follows from equations (4), (5), (6), and (42) that we can express

J , u, Y , and w as functions of x and non-policy parameters. It then follows that j ≡ J/Y is a

function of x. Second, a firm that sets its price in period 0 will set the optimal markup over current

marginal cost because all prices will be re-optimized in subsequent periods. Therefore the optimal

relative price for a firm that updates in period 0 is µ(w0 + ψ1M

ψ20 /h(x0, τ)

)/A0 where h(x, τ) is

given by (20). The price level and inflation rate in period 0 are therefore functions of x0 and τ .

Third, it follows from (8) that the same is true of the nominal interest rate I0 and S0. The Fisher

equation (46) and the assumption that the central bank sets inflation to zero in period 1 implies

that R0 is also a function of x0 and τ .

The next step is to derive an “aggregate demand curve” from the aggregate Euler equation.

Using equations (17), (18), and (19) along with the log-normal income process we arrive at:

(1− ut + utb)

Ct= βRte

σ2ε (xt)(1−τ)

2

Et

[(1− ut+1(1− b))

(1− ut+1(1− b−1)

)Ct+1

].

For t ≥ 0 the expectation in the Euler equation disappears because outcomes in t + 1 are deter-

ministic. We can re-write the Euler equation as

C0 =1− u0 + u0b

βR0(x0)

C1

(1− u1(1− b)) (1− u1(1− b−1))e−σ

2ε (x0)(1−τ)

2

.

Next, recall from market clearing that: Y0(1− j0) = C0(1 + χηG0 ) Replacing out C0 gives:

Y0 =1 + χηG01 + χηG1

1− j11− j0

1− u0 + u0b

βR(x0)

Y1(1− u1(1− b)) (1− u1(1− b−1))

e−σ2ε (M0)(1−τ)2 . (58)

This provides an “AD” curve: a negative relation between x0 and Y0.

The AD curve depends on Y1. Under flexible prices, the price-setting first-order condition

56

implies

w(xt) +ψ1x

ψ2t

h(xt, τ)=ηAtµ.

This equation implies xt is a function of ηAt , τ , and non-policy parameters when prices are flexible.

We call this function xFlex(ηAt , τ). Y1 can be calculated from (42) with x1 = xFlex(ηA1 , τ) because

all prices are flexible in period 1. This leads to

Y1 = ηA1 (1− u(x1)) [w(1− τ)]1

1+γ (x1)ζ

1+γ (59)

Note that x1 is given by xFlex(ηA1 , τ) so (59) makes Y1 independent of x0 and b.

The aggregate production function, (42), is the “AS” curve: a positive relation between x0 and

Y0. We now set AS = AD. Specifically, equations (58) and (42) lead to two expressions for Y0.

Taking the log of equations (58) and (42) and set them equal to one another yields an implicit

solution for x0:

log

(1− χηG01− χηG1

)+ log(1− j1)− log(1− j0) (60)

+ log (1− u0 + u0b)− log β − log (R(x0, τ)) + log(Y1(x

Flex1 , τ)

)− log (1− u1(1− b))

− log(1− u1(1− b−1)

)− σ2ε (x0) (1− τ)2 − log

(ηA0)

+ log (S0)− log (h0)− log ((1− u0)) = 0.

D.4 Proof of proposition 4

x0 is implicitly defined by (60). Using the implicit function theorem we obtain

dx0db

= −u0

1−u0+u0b −u1

1−u1(1−b) + u1b−2

1−u1(1−b−1)

− 1−b1−u0+u0b

du0dx0− 1

R0

dR0dx0− (1− τ)2 d

dx0σ2ε (x0) + 1

S0

dS0dx0− 1

h0dh0dx0

+ 11−u0

du0dx0

+ 11−j0

dj0dx0

As an elasticity we have

b

x0

dx0db

=

u0b1−u0+u0b −

u1b1−u1(1−b) + u1b−1

1−u1(1−b−1)

d logR0

d log x0+ (1− τ)2σ2ε (x0)

d log σ2ε (x0)

d log x0+ 1−b

1−u0+u0bu0d log u0d log x0

− d logS0

d log x0+ d log h0

d log x0+ d log(1−u0)

d log x0+ d log(1−j0)

d log x0

(61)

57

D.5 Proof that the numerator of (33) is increasing in u0 and u1 for u < 1/2

Start with u0b1−u0(1−b) , the derivative with respect to u0 is

b(1− u0(1− b)) + u0b(1− b)[1− u0(1− b)]2

=b

[1− u0(1− b)]2> 0.

Next turn to bu1(1−u1)(b−2−1)[1+u1(b−1−1)][1+u1(b−1)] . We start by dividing by b(b−2 − 1) ,which is positive for

b ∈ (0, 1). We will then show the derivative with respect to u1 is positive

1− 2u1

[1 + u1(b−1 − 1)] [1 + u1(b− 1)]2[1 + u1(b

−1 − 1)]

[1 + u1(b− 1)]

− u1(1− u1)[1 + u1(b−1 − 1)] [1 + u1(b− 1)]2

[(b−1 − 1

)(1 + u1(b− 1)) + (b− 1)

(1 + u1(b

−1 − 1))]> 0

multiply both sides by[

1 + u1(b−1 − 1)

][1 + u1(b− 1)]

2[1− 2u1]

[1 + u1(b

−1 − 1)]

[1 + u1(b− 1)]− u1(1− u1)[(b−1 − 1

)(1 + u1(b− 1)) + (b− 1)

(1 + u1(b

−1 − 1))]> 0

Rearranging

(1− 2u1)[1 + (u1 − u21)(b−1 + b− 2)

]− u1(1− u1)

(b−1 + b− 2

)(1− 2u1) > 0

divide both sides (1− 2u1) (recall u1 < 1/2) and rearrange

1 + u1(1− u1)(b−1 + b− 2)− u1(1− u1)(b−1 + b− 2

)= 1 > 0.

D.6 Proof of proposition 5

Using the implicit function theorem to differentiate (60) we obtain

dx0dτ

= −2σ2ε (x0) (1− τ) + 1

Y11

1+γY11−τ −

1h0

11+γ

h01−τ + 1

S0

∂S0∂τ

∣∣∣x− 1

R0

∂R0∂τ

∣∣∣x

+ 1Y1

dY1dx1

dxFlex(A1,τ)dτ

− 1−b1−u0+u0b

du0dx0− 1

R0

dR0dx0− (1− τ)2 d

dx0σ2ε (x0) + 1

S0

dS0dx0− 1

h0dh0dx0

+ 11−u0

du0dx0

+ 11−j0

dj0dx0

= −2σ2ε (x0) (1− τ)− 1

R0

∂R0∂τ

∣∣∣x

+ 1S0

∂S0∂τ

∣∣∣x

+ 1Y1

dY1dx1

dxFlex(A1,τ)dτ

− 1−b1−u0+u0b

du0dx0− 1

R0

dR0dx0− (1− τ)2 d

dx0σ2ε (x0) + 1

S0

dS0dx0− 1

h0dh0dx0

+ 11−u0

du0dx0

+ 11−j0

dj0dx0

58

As an elasticity we have

τ

x0

dx0dτ

=2σ2ε (x0) (1− τ) τ − ∂ logR0

∂ log τ

∣∣∣x

+ ∂ logS0

∂ log τ

∣∣∣x

+ d log Y1d log x1

d log x1d log τ

d logR0

d log x0+ (1− τ)2σ2ε (x0)

d log σ2ε (x0)

d log x0+ 1−b

1−u0+u0bu0d log u0d log x0

− d logS0

d log x0+ d log h0

d log x0+ d log(1−u0)

d log x0+ d log(1−j0)

d log x0

(62)

D.7 Proof of lemma 3

See (61) and (62).

E Description of methods for section 6

E.1 Estimated income process

The material in this appendix describes an implementation of the procedure of Guvenen and McKay

(2016) and that paper gives more discussion and details.

The income process is as follows: αi,t evolves as in (3). Earnings are given by αi,twt when

employed and zero when unemployed. Notice that here we normalize ht = 1 and subsume all

movements in ht into wt. While this gives a different interpretation to wt it does not affect the

distribution of earnings growth rates apart from a constant term. The innovation distribution is

given by

εi,t+1 ∼ F (ε;xt) =

N(µ1,t, σ1) with prob. P1,

N(µ2,t, σ2) with prob. P2,

N(µ3,t, σ3) with prob. P3

N(µ4,t, σ4) with prob. P4

59

The tails of F move over time as driven by the latent variable xt such that

µ1,t = µt,

µ2,t = µt + µ2 − xt,

µ3,t = µt + µ3 − xt,

µ4,t = µt.

where µt is a normalization such that Ei[expεi,t+1] = 1 in all periods.

The model period is one quarter. The parameters are selected to match the median earnings

growth, the dispersion in the right tail (P90 - P50), and the dispersion in the left-tail (P50-P10) for

one, three, and five year earnings growth rates computed each year using data from 1978 to 2011.

In addition we target the kurtosis of one-year and five year earnings growth rates and the increase

in cross-sectional variance over the life-cycle. The moments are computed from the Social Security

Administration earnings data as reported by Guvenen et al. (2014) and Guvenen et al. (2015). We

use these moments to form an objective function as described in Guvenen and McKay (2016).

The estimation procedure simulates quarterly data using the observed job-finding and -separation

rates and then aggregates to annual income and computes these moments. To simulate the income

process, we require time series for xt and wt. We assume that these series are linearly related to

observable labor market indicators (for details see Guvenen and McKay, 2016). Call the weights

in these linear relationships β. We then search over the parameters P , µ, σ, and β subject to the

restrictions P2 = P3 and σ2 = σ3.

Guvenen et al. (2014) emphasize the pro-cyclicality in the skewness of earnings growth rates.

The estimated income process does an excellent job capturing this as shown in the top panel

of figure 5. The estimated β implies a time-series for xt which shifts the tails of the earnings

distribution and gives rise the pro-cyclical skewness shown in figure 5. We regress this time-series

on the unemployment rate and find a coefficient of 16.7.31 The fourth component of the mixture

distribution occurs with very low probability, and in our baseline specification we set it to zero.

This choice is not innocuous, however, because the standard deviation σ4 is estimated to be very

large and this contributes to the high kurtosis of the earnings growth distribution. In particular,

31 We regress this estimated time series xt on the unemployment rate, which we smooth with an HP filter withsmoothing parameter 100,000. If we call this regression function f , we then proceed with F (ε′; f(u)).

60

omitting this component leads to a substantially smaller τ as a result of having less risk in the

economy. We prefer to omit this from our baseline calibration because the interpretation of these

high-kurtosis terms is unclear and we are not entirely satisfied with modeling them as permanent

shocks to skill.

The resulting income process that we use in our computations is as follows: The innovation

distribution is given by

εi,t+1 ∼ F (ε;xt) =∼

N(µ1,t, 0.0403) with prob. 0.9855,

N(µ2,t, 0.0966) with prob. 0.00727,

N(µ3,t, 0.0966) with prob. 0.00727

with

µ1,t = µt,

µ2,t = µt + 0.266− 16.73(ut − u∗),

µ3,t = µt − 0.184− 16.73(ut − u∗),

where u∗ is the steady state is unemployment rate in our baseline calibration. The bottom panels

of figure 5 show the density of ε and how it changes with an increase in the unemployment rate.

E.2 Global solution method

As a first step, we need to rewrite the Calvo-pricing first-order condition recursively:

p∗tpt

=Et∑∞

s=tR−1t,s (1− θ)s−t

(ptps

)µ/(1−µ)Ysµ

(ws + ψ1M

ψ2s /hs

)/As

Et∑∞

s=tR−1t,s (1− θ)s−t

(ptps

)1/(1−µ)Ys

.

Define pAt as

pAt = Et∞∑s=t

R−1t,s (1− θ)s−t(ptps

)µ/(1−µ)Ysµ

(ws + ψ1M

ψ2s /hs

)/As

61

1975 1980 1985 1990 1995 2000 2005 2010-0.2

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

0.2Kelley's skewness

Figure 5: Properties of F (ε).62

and pBt as

pBt = Et∞∑s=t

R−1t,s (1− θ)s−t(ptps

)1/(1−µ)Ys

such that

p∗tpt

=pAtpBt.

pAt and pBt can be rewritten as

pAt = µYt

(wt + ψ1M

ψ2s /hs

)/At + (1− θ)Et

[(Itπt+1

)−1π−µ/(1−µ)t+1 pAt+1

](63)

pBt = Yt + (1− θ)Et

[(Itπt+1

)−1π−1/(1−µ)t+1 pBt+1

]. (64)

The procedure we use builds on the method proposed by Maliar and Maliar (2015) and their

application to solving a New Keynesian model. We first describe how we solve the model for a

given grid of aggregate state variables and then describe how we construct the grid.

There are six state variables that evolve according to

Ei[α1−τi,t+1

]= (1− δ)Ei

[α1−τi,t

]Ei[ε1−τi,t+1ut

]+ δ

Ei [logαi,t+1] = (1− δ) [Ei [logαi,t] + Ei [log εi, t+ 1|ut]]

SAt+1 = St

log ηAt+1 = ρA log ηAt + εAt+1

log ηGt+1 = ρG log ηGt + εGt+1

log ηIt+1 = ρI log ηIt + εIt+1,

where SA is the level of price dispersion in the previous period and the ε terms are i.i.d. normal

innovations.

There are five variables that we approximate with complete second-order polynomials in the

state: (1/Ct), pAt , pBt , Jt, and Vt, where Vt is the value of the social welfare function. We use (17)

and (18) to write the Euler equation in terms of Ct and this equation pins down 1/Ct. pAt and pBt

63

satisfy (63) and (64). Vt satisfies

Vt = Wt + β Et [Vt+1] .

Jt satisfies Jt = ψ1Mψ2t (υ− ut). Abusing language slightly, we will refer to these variables that we

approximate with polynomials as “forward-looking variables.”

The remaining variables in the equilibrium definition can be calculated from the remaining

equations and all of which only involve variables dated t. We call these the “static” variables.

To summarize, let St be the state variables, Xt be the forward-looking variables, and Yt be the

static variables. The three blocks of equations are

S ′ = GS(S,X ,Y, ε′)

X = EGX (S,X ,Y,S ′,X ′,Y ′)

Y = GY(S,X )

where GS are the state-transition equations, GX are the forward-looking equations and GY are the

state equations. Let X ≈ F(S,Ω) be the approximated solution for the forward-looking equations

for which we use a complete second-order polynomial with coefficients given by Ω. We then opera-

tionalize the equations as follows: given a value for S, we calculate X = F(S,Ω) and Y = GY(S,X ).

We then take an expectation over ε′ using Gaussian quadrature. For each value of ε′ in the quadra-

ture grid, we compute S ′ = GS(S,X ,Y, ε′), X ′ = F(S ′,Ω) and Y ′ = GY(S ′,X ′). We can now

evaluate GX (S,X ,Y,S ′,X ′,Y ′) for this value of ε′ and looping over all the values in the quadrature

grid we can compute X = EGX (S,X ,Y,S ′,X ′,Y ′). X will differ from the value of X that was

obtained initially from F(S,Ω). We repeat these steps for all the values of S in our grid for the

aggregate state space. We then adjust the coefficients Ω part of the way towards those implied by

the solutions X . We then iterate this procedure to convergence of Ω.

Evaluating some of the equations of the model involves taking integrals against the distribution

of idiosyncratic skill risk εi,t+1 ∼ F (εi,t+1, ut − u∗). We do this using Gaussian quadrature within

each of the components of the mixture distribution. We compute expectations over aggregate

shocks using Gaussian quadrature as well.

We use a two-step procedure to construct the grid on the aggregate state space. We have six

aggregate states so we choose the grid to lie in the region of the aggregate state space that is visited

64

by simulations of the solution. We create a box of policy parameters [bL, bH ] × [τL, τH ]. For each

of the four corners of this box, we use the procedure of Maliar and Maliar (2015) to construct a

grid and solve the model. This procedure iterates between solving the model and simulating the

solution and constructing a grid in the part of the state space visited by the simulation. This gives

us four grids, which we then merge and eliminate nearby points using the techniques of Maliar

and Maliar (2015). This leaves us with one grid that we use to solve the model when we evaluate

policies. Each of the grids that we construct have 100 points.

65